20VC: Why AI Cannot Replace Humans in Enterprise | Why Work Processes Not Models Will Be The Most Valuable Asset in AI | Why Europe Has Lost and Building in the US vs EU with Daniel Dines, UiPath
- Daniel Dines argues that today’s AI can remember work without being transformed by it, making “millions of Einsteins in a data center” reasoning engines rather than hireable people. Humans alter through experience; models append context without changing their weights on the job. Until models develop human-like initiative, will, and experiential learning, Dines expects enterprise adoption to diffuse process by process—not erase employment overnight.
- The durable enterprise asset is not the model but the “map of work”: the workflows, exceptions, systems, relationships, and unwritten judgment that make a company function. Models are interchangeable, so enterprises need portable documentation that lets them switch providers, train internal models, and preserve their IP. “The workflow, the map of work and the workflows around the map of work is where the real value is.”
- AI’s probabilistic nature makes it the design layer, while deterministic software remains the execution layer for consequential work. Dines illustrates that a 99%-reliable step repeated 100 times leaves only roughly a 60% chance of completing the whole sequence. He therefore sees coding agents generating, repairing, testing, and auditing exact software rather than autonomously improvising through production. “You use AI to create software that runs the enterprise in a predictable, governed, auditable way.”
- Workforces will shrink in many functions, but blunt AI layoffs risk removing precisely the initiative, trust, and institutional memory needed to deploy AI successfully. Dines expects fewer credentialed specialists yet greater value for employees who manage exceptions, mentor others, maintain customer relationships, and exercise judgment outside the formal job description. His prescription is a workforce ledger that captures those hidden outputs before deciding “which ones” to retain or redeploy.
- Vibe coding has collapsed prototyping costs without eliminating the expensive path from prototype to production. UiPath built a procurement tool only with AI, then found missing connectors, permissions, audits, security requirements, insufficient tests, and a “completely bogus” database schema that required human intervention. Dines’s warning: internal replacements may eventually cost as much as purchased software while consuming the company’s best technical bandwidth.
- Enterprise model traffic should concentrate in cheap models, while strategic value accrues to workflow owners, open-model infrastructure, and contextual data systems. Dines predicts 90% of operational flow will use cost-efficient models, with responsible enterprises maintaining an open-source fallback. He would “probably” invest in Fireworks at $15 billion if it can secure the compute needed for scaled inference, potentially requiring tens of billions in capital. In legal AI, the difference between a model call and a valuable company is whether it maps and operates the full legal workflow. He also argues Jensen is “bound by the success of open source,” since a closed frontier duopoly could eventually make its own chips.
- Dines sees Europe as technologically “largely irrelevant” despite possessing the talent and chipmaking machinery, because US companies make faster, larger bets. He advises builders of universal technology to move to America, while identifying European demand for on-prem software, model sovereignty, and optionality as a real countervailing market. UiPath’s bull case rests on supplying the “map and rails” beneath enterprise agents; its bear case is genuine, near-free “Einsteins” capable of replacing whole people.
1. Models can reason like Einstein without becoming Einstein
Dines wrote his book partly to order his own thinking, using Claude and ChatGPT as “ghostwriters” through an almost six-month process. The animating question was whether durable AI limitations remain—or whether millions of digital Einsteins soon perform all economically useful work while humans “go to play.”
His distinction is between possessing some of Einstein’s reasoning power and being an Einstein-like person who learns through work. A chef formed by 20 years of Japanese cooking will interpret the same recipe differently from one formed by Italian cooking; reading every chess book does not make someone a grandmaster, just as watching skiing videos does not make someone a skier.
Harry’s pushback—worth keeping—is that models already retain memories, avoid fatigue, demand no raises, and can execute much of finance, marketing, sales, and social media. Dines’s rebuttal is that a scratchpad changes the prompt, not the model: “Memory, it’s not necessarily learning.” Humans carry conversations forward as transformed people; deployed models retain the same weights.
Recursive self-improvement might eventually alter that constraint, but Dines separates reasoning from will. Infinite compute could conceivably yield simulations as complex as the world, yet he calls it “wishful thinking” that a sufficiently large self-improving model must generate will. AI solves novel math problems, he notes, but still does not create frameworks comparable to relativity; style and individuality likewise require being transformed by experience.
2. “Pacing the frontier” may conceal a fight over open source
Dines’s safety standard is direct: if frontier labs truly believe their experiments could become uncontrollable and cause material harm, they should slow down “at any cost,” without waiting for government pressure. A concerned builder should already fear legal responsibility—and, in his blunt formulation, going to jail.
He interprets calls to coordinate the “good guys” as implicitly seeking freedom to continue building while limiting consequences. Unknown bad actors cannot realistically be persuaded into a global pause; Dines would classify even today’s Chinese AI labs as good actors, leaving uncontrolled downstream access as the argument’s actual target.
That makes the safety memo, in his reading, “indirectly…an attack on open source”: even benevolent developers can release capabilities that reach malicious users. The interpretation is hedged rather than asserted, but it matters because open models are also the primary hedge against concentration among frontier providers.
Large enterprises are cautious, though less because OpenAI might manufacture screws than because proprietary information might leak into intelligence available to existing competitors. Dines calls that a legitimate concern: companies need to protect their IP and retain a verifiable path away from any one closed model.
3. AI should create exact software, not improvise every transaction
Dines’s second durable limitation is “exactness.” A probabilistic agent that is 99% reliable at each step has, in his illustration, roughly a 60% chance of completing 100 steps correctly; across hundreds or millions of operations, small error rates compound. Capability therefore does not imply suitability: a model can multiply numbers, but a computer remains the correct execution engine.
Harry suggests convenience wins because users ask whatever environment they already inhabit. Dines agrees at the interface layer: ChatGPT translates natural language into a tool call, while deterministic computation supplies the answer. The enterprise analogue is to route every task requiring exactness onto technology that behaves identically for a given input.
This produces an asymmetry: deploying reliable autonomous agents is not getting easier than it was two years ago, but creating automation has become dramatically easier. Dines ranks coding agents alongside ChatGPT and chain of thought as major milestones because they operate at design time, generating deterministic systems that execute repeatedly without changing behavior in production.
AI can also fix an automation when an upstream system changes. Humans can audit the generated software, validate it, and build tests guaranteeing behavior. Dines’s emerging pattern is therefore not a probabilistic model directly running the enterprise, but AI “creating the software that runs an enterprise” inside predictable, governed rails.
4. AI workforce planning starts with invisible human outputs
UiPath has roughly 4,000 employees, including more than 1,000 engineers. Dines has told them transformation is unavoidable, but rejects using AI as a pretext for arbitrary cuts: workforce restructuring should happen alongside successful enterprise adoption, not as a 20% RIF followed by a promise that automation will eventually justify it.
Every job produces measurable output plus less legible institutional value: customer trust, mentorship, cultural continuity, initiative, or the hunch that an account may churn before data confirms it. Cutting roles because agents can send emails may destroy the relationship that kept the customer. Dines wants an enterprise ledger recording these secondary outputs before headcount decisions.
His “credentialed middle” may be especially exposed: companies historically hired credentialed domain expertise, precisely the knowledge AI can now supply broadly. Yet fewer experts does not mean fewer valuable humans; initiative, AI literacy, exception handling, and the ability to maintain relationships may matter more than narrow expertise during the transition.
Harry’s trainee-lawyer example sharpens the contraction: a program that historically hired 25 expects to take four. Dines agrees most roles may need fewer people but contests Harry’s verifiability test. The operative variable is whether someone else has defined the frame; even invoices contain undocumented customer priorities, making the selection problem “which ones?” rather than whether numbers reconcile.
5. Cartography turns tacit work into a deployable map
Dines defines the “map of work” as every workflow, exception, procedure, and system used to accomplish a process. An enterprise cannot hand AI a neat job description and expect competence; it must surface the unwritten choices accumulated by employees who have lived the work.
UiPath’s proposed discipline is “cartography.” Its Cartographer Agent observes subject-matter experts at their desktops, records their work, and interviews them in real time: why did a different ZIP code change the invoice path, and why was this exception handled differently? Evidence from multiple employees is consolidated into an as-is process map.
From that map, coding agents can redesign the process and “print” the required software. The desired transition is from people manually operating systems of record toward automation and agentic AI operating more of those systems—with humans concentrated where exceptions, initiative, relationships, and accountability remain essential.
Harry raises the obvious employee reaction: observation can look like surveillance designed to automate them away. Dines calls the transformation inevitable but says trust depends on the message—no mass extinction under the pretense of AI, a genuine chance for employees to become AI-literate, and adoption one process at a time. Fear has eased as workers see that the Einsteins are “not hireable yet.”
6. Replacement economics work only after production reality
Dines would hire a machine even at greater cost if it genuinely matched or exceeded a human: labor costs rise and introduce errors, while machine costs should decline, creating a future advantage. He therefore treats inference expense as secondary to capability. The present blocker is simpler: “This Einstein doesn’t yet exist.”
Harry counters with Jason Lemkin reducing a team from 25 to two and claiming AI replaced finance and marketing leadership. Dines refuses to extrapolate from a single founder or narrow setting, citing companies that removed hundreds of support workers and later rehired them. Replacement must be demonstrated at scale across industries before becoming a general rule.
UiPath’s own vibe-coding experiment initially looked extraordinary. A procurement application was written only by AI, but production exposed missing connectors, permissions, audits, security, insufficient testing, and a “completely bogus” database schema. Business users still could not own the full lifecycle without engineers and other people restructuring and maintaining the system.
That is why Dines would still buy Salesforce as a system of record. Prototypes are now exceptionally cheap; production remains where the work lives. An internal substitute can end up costing as much or more than the software it replaces while tying up the company’s strongest people.
7. Workflows, cheap models, and contextual data capture the value
Legal AI illustrates both disruption and revenue compression. Harry sizes US legal services at $300 billion and suggests 30% automation creates $90 billion of opportunity; Dines argues that perhaps 10% of that becomes token revenue. A generic legal opinion can come from interchangeable frontier or open models—the larger prize belongs to products that map workflows and effectively operate a legal department.
Dines predicts 90% of enterprise operational traffic will go to highly cost-efficient models rather than true frontier systems such as “Astra or Fable,” as spoken. Anthropic and OpenAI may serve much of it through cheaper models, but responsible enterprises should maintain a verifiable open-source backup and preserve the ability to switch.
The map of work becomes the company’s core AI IP: the documented “who am I” needed to train today’s internal model and transfer knowledge to a better base model two months later. Dines expects enterprises at least to maintain their own models as backup, while acknowledging frontier providers may still deliver more intelligence per dollar through infrastructure scale.
He would “probably” invest in Fireworks at $15 billion if the open-model thesis holds, but says it must secure compute—potentially requiring tens of billions in capital—to provide inference at the desired scale. On data providers, he distinguishes storage from intelligence: raw data is tape; value lies in selecting the right context and feeding it to the model at the right moment.
The same open-model thesis underpins his view of NVIDIA: if OpenAI and Anthropic became a duopoly, they could eventually make their own chips, so Jensen is “bound by the success of open source.” He sees NVIDIA’s support for open-source ecosystems as aligned with protecting its infrastructure opportunity.
8. UiPath’s bet is “map and rails” in a US-led AI economy
Dines calls Europe technologically “largely irrelevant,” despite ASML, abundant talent, and European roots among leading AI builders. The problem is commercial culture: US companies make larger vision-led bets with fewer proof points, and even middle managers can authorize million-dollar experiments. His advice to a young European building universal technology is therefore, reluctantly, “go to the US.”
Europe still offers a sovereignty wedge. Customers prefer on-prem software, model sovereignty, and model optionality; Dines has urged Fireworks to offer an on-prem product. Unlike the US “show me the money” posture, European buyers first need to see proven technology—but he believes meaningful business follows.
UiPath reported roughly $1.6 billion in revenue and 14% growth for the prior year. Harry says public markets punish companies growing below 20%; Dines agrees that markets driven by sentiment can automatically classify slower-growing software companies as AI losers. He also argues that many 2021 private-company “zombies” might fare better in public markets because employees and investors would at least have a route to liquidity, even if the market exposed their real valuations.
The bull case for a $50 billion UiPath rests on its evolution from RPA into orchestration. Dines points to Gartner’s new BOOT—Business Orchestration and Automation Technologies—Magic Quadrant, where UiPath moved from challenger to leader, as well as positive views from Gartner, Forrester, and other analysts. Agents receive a goal, a map describing reality, and rails constraining permissible actions.
On infrastructure, Dines says every major infrastructure cycle is overbuilt: participants may be building 200% of a 100% opportunity. He does not think AI infrastructure is overbuilt for the next decade, but it could be overbuilt for the next three years if human-work replacement takes ten; timing can be merciless.
The bear case is that models become genuine people-like Einsteins, token costs approach zero, and enterprises can assign them any job without needing the same maps or deterministic rails. Harry cites token prices falling from $60 to $1 per million; Dines concedes costs may vanish, but says the real question is whether the capability to replace a person arrives. Meanwhile, he spends about half his day in Visual Studio Code with Claude and ChatGPT, using a strategy folder and agents to multiply his leverage.
Full transcript
In my opinion, even the labs in China that are building AI right now, I would classify them as the good guys. I've never hidden from my employees that there will be a transformation. Jensen is bound by the success of open source. Models are interchangeable, but the workflow, the map of work, and the workflows around the map of work are where the real value is.
Daniel, dude, it is so good to have you in the hot seat. I've been looking forward to this one, so thank you so much for joining me, dude.
Likewise, dude. It's always a pleasure to be here, and I think it's the hottest moment in technology, so I'm very excited to talk to you and get your perspective on a lot of topics.
It really is the most wild time right now, and so I want to start. You've written a book. A lot of people write books. With the greatest of respect, you are my friend, and I care deeply about you: why on earth did you decide to write a book as a public company CEO? No offense. You're not doing it for the royalties. Why did you decide to write a book?
Well, I dreamt of writing a book since I was a kid, and I discovered I had no talent. I got a really great opportunity. Claude and ChatGPT helped me a lot; they were my ghostwriters. It was a good moment to put my own ideas in order, because when you write something, you get much clearer perspectives on what you are doing.
It was almost a 6-month effort, and I started with different threads of thought. One was: what are the limitations of AI? Is there any durable limitation of AI where, in a couple of years, there will be millions of Einsteins in a data center, and we can all go play or do whatever we would like to do because the Einsteins will do the work for us?
Can we start on that, then? I think it's nice to take it in segments. Limitations of AI, Einsteins in data centers: how should we think about that moving forward?
1. AI Cannot Learn On The Job
When I heard the statement that, in a couple of years, we will have millions of Einsteins in a data center, I was really concerned. Dario is someone I highly respect, and he's highly successful. So I was thinking, "What does it mean for us? What does it mean for me? Can I hire one of these Einsteins, put it in a laptop somehow, assign an enterprise account and a Slack account, and ask it to do my job or whatever job in the enterprise?"
The reality may be different. Probably Dario wanted to say that we'll have millions of entities that will have some of the reasoning powers of Einstein, which I agree with, but not Einsteins as people—not Einsteins that are capable of learning on the job. Do you agree that one of the major expectations when you hire someone is that they will learn on the job? There is no manual that a company can give a new employee saying, "This is exactly how you do your job from end to end." So we expect that—
No, I think humans do learn on the job, and they improve on the job, as do models. It's similar there, except humans get tired. Humans want more money. Humans want culture. Humans can be toxic. Humans are difficult to manage. I'll take the AI any day of the week, please.
If AI can work as well as a human, Harry. But you say that AI learns on the job. AI can create a notepad on the job, a scratchpad where it can memorize some of the policies on the job. But AI doesn't alter its weights on the job in the way humans are transformed by a job. This is a huge difference.
Let me give you an example. You have 2 chefs. One chef has spent 20 years doing only Japanese food, and the other has done only Italian food, and you give them 1 recipe. They will create different food. It's not that you can write down your enterprise on a sheet of paper. It's much more complex. It's a becoming.
Think about it this way: if I give someone the ability to read all the books about chess, do you think they'll become a grandmaster without playing, without losing, without going through all of this process? Probably not. If I have you watch all the videos about skiing, are you becoming a skier? No, you are not becoming a skier.
But I think it depends on what task and workflow you're doing within the enterprise. If we look at the majority of what people within UiPath and every company do—whether it's accounting and finance, marketing and sales, largely outbound and inbound, or social media—most of this is execution-oriented. Yes, judgment, ambiguity, and taste are important at the top, but most of what people do is execution.
I disagree with you. I think most people display some sort of micro-initiatives during the job. Maybe I have a hunch that this customer is going to churn, and I can act before any data is even available. How do I develop this hunch? It's through my years of transformation. It's not written on a piece of paper.
Harry, it's a big difference between writing an operating model on a piece of paper and living it. It's almost impossible for an enterprise. For AI, do you admit that everything has to be written down and documented? Every time I ask a question of the model, the model will have to read my entire enterprise.
Yep, sure.
So this is not possible.
But I also think you're talking about today's state of AI.
This is one of the biggest bottlenecks right now, because AI doesn't train on the job or train its own weights. Every time I do something, after this talk with you, I am being transformed. I carry this discussion with me in all my thoughts. This is not true about AI.
It is. That's why people remain with OpenAI: it has memory, and it is able to infer from past queries and prompts and give you suggestions based on those. So it does have memory.
It has memory, but memory is not necessarily learning. It's not the same thing. Memory is just something that is written down. When I'm talking to you, I don't go back into my memory. I am just being transformed.
It's like a model—like a new version of the model that comes improved. The new version is not the old version plus a piece of paper that has been memorized. It's transformed in its own weights. This is why a model becomes much better.
I want to give you a simple example using our own technology. We make our UiPath platform available to coding agents, so it's much easier to create automations on UiPath right now. But what we discovered is that a model that has read open-source technology, has a lot of examples, and already has a certain technology in its weights will be much better than creating on our own technology.
Because regardless of how many prompts and skills we create, the model has it in its own weights. It's very different. Think of all the metaphors in the world when you read something versus when you live something. You can read a biography; it doesn't mean you live that life. It doesn't mean you are transformed and you're going to answer like the person who lived it. To me, this is really the biggest limitation that the models have right now.
So I actually do agree with you, but I think everyone does. That's why everyone is chasing recursive self-improvement so much: models that can continuously learn from themselves and improve over time without the need for human intervention. Does that not remove the limitation that we just discussed?
I don't know, man. Maybe we are the result of a self-improvement loop. Let's do this thought exercise: Let's put a model that we have today, with the best technology, in a spaceship and throw it to the stars. Let's say that we have this technology, like I think von Neumann imagined, that is self-replicating. This spaceship goes to different stars, gets energy, and can continue. Compute will be infinite, and models will self-improve. Where would they end up?
Maybe they would create a simulation of a world like ours, right? Because they would improve infinitely, basically. This is the theory. So they would simulate a world as complex as our own world within it. But that means that we are part of an infinite simulation. I don't know where it's going to lead, but I know that there is a big distinction that I made in the book between will and reasoning.
It's not as if we are certain that the will to do something emerges from reasoning or even from consciousness. I think will is a separate part of the fabric of the universe. I don't think we, as humanity, have clarity about what will is. I think it's wishful thinking to believe that I can take a big model, put it into a self-improvement loop, and that this model is going to generate will. I don't believe so.
But I think we don't know, and I think that's what's so challenging about trying to predict what happens. It's a world of, as I said, recursive self-improvement. A technology you create can become something you didn't know it could be. I guess for me, the question then is: We see the news this week—Dario says we need to pace the frontier. You run UiPath today. Do you feel we need to pause the frontier?
2. Frontier Labs Face A Safety Choice
I would say that if they truly believe this technology is becoming rogue and they cannot control it, and their experiments will create significant loss for the internet or other systems, if I were them, I would pace it at any cost because I don't want to risk going to jail, honestly. I think there are laws that control this type of rogue behavior. So honestly, I don't need external pressure to control it. I would just be a concerned citizen, and I would not build a technology that is causing harm.
Now, of course, they think that this is the only way to protect against the bad boys. So we are the good guys, but there will be some bad boys, probably in other parts of the world, who will build the technology regardless. So I think they are probably asking more for a pause. They want to build this technology at any risk, and they're willing to open their gates for others to see how they're doing it because they want to do it in as good a manner as possible. But at the same time, they want to be free of consequences.
To me, I think this is a bit how I read this memo, because otherwise I think it's kind of obvious. I don't think we can reason with the bad guys or make a coalition with the bad guys to stop the frontier. So we can make a coalition only with the good guys, regardless. In my opinion, I would even take the labs in China that are building AI right now and classify them as the good guys.
To me, I think the indirect attack is probably on open source, because they are basically saying that even if the good guys are building open source, that open source will get into the hands of the bad guys. These are unknown bad guys. This is the real danger. So the danger is in open source. Indirectly, it's also an attack on open source in this way. It's a way of interpreting it, I guess.
You work with some of the biggest enterprises in the world with UiPath. Alex Karp from Palantir said that the biggest enterprises in the world are scared to work with frontier labs because of the threat of them coming into their businesses over time. They have the data. They could build their own and compete against them. Do you see large enterprises being scared to work with frontier providers?
I think so, yes. I don't think people are scared that OpenAI will build a competitor to them, necessarily. I don't see this coming. They are more scared that their IP would leak to other existing competitors somehow.
I don't know—if I'm manufacturing screws or whatever, I don't think OpenAI is going to come and compete with me on this thing. But probably some of the other guys can get, indirectly through other models, the same intelligence if OpenAI trains the model. I think that's the real danger, and I think it's a legitimate danger. Everyone is trying to protect their IP.
You said that writing leads to a clear articulation of thoughts. What was another thought that you clearly articulated through the writing process?
3. Exactness Powers Enterprise Automation
It has become very clear to me that another limitation of AI is what I call exactness. AI, by its nature of being probabilistic at every step, can lose it. When you do 100 or 200 steps with AI, even if at each step you have a 99% probability, for instance, it's 0.99 to the power of 100. You will end up with maybe a 60% probability of completing the entire step.
AI doesn't have the mechanism to follow steps exactly hundreds or millions of times in the same way. You can see it even if you ask AI to multiply very large numbers millions of times. At some point, it will make a mistake. It also surfaces another simple idea: Even if you have a tool like AI that is capable of doing multiplications, why are you not using a computer that can do these multiplications millions of times, 100% of the time, exactly? The fact that the tool can do a job doesn't mean you have to use that tool to do that type of job.
Isn't it because it's where you are? That's the importance of being in the harness of the workflow. You could use something else. I'm asking here; I'm not saying. But because you're in ChatGPT continuously every day, instead of switching to a computer or calculator or whatever, you ask what you're already in. That's the importance of the environment.
ChatGPT acts as an interface to convert my questions in natural language into exactness, but the exactness is not run by ChatGPT. Exactness is run by a computer because even today, if you ask ChatGPT, “Please multiply these 2 big numbers,” it uses a tool—a computer—behind the scenes and gives you the exact number. This is part of the power of the models.
You can see, at the desktop level, in this type of co-work or ChatGPT work, the capability to call tools. What I am saying is that you extend this capability to the enterprise level, where everything that should be exact should run on exact technologies. There is no point in running it on probabilistic technologies.
Here comes, I think, the most interesting thing: There isn't a symmetry in the deployment of AI and automation in an enterprise. Deploying AI agents is not getting easier today than it was 2 years ago, in my opinion. But deploying automation has become much easier because I can create these automations with AI, with coding agents.
Coding agents have been the major giant leap that we have seen in the past year. I would say that since the invention of ChatGPT, chain of thought and then coding agents were the major milestones. With coding agents that act at design time, when I build the systems, I can create automations that work with exactness every time during execution time.
This is the asymmetry that is happening right now. Plus, when an automation breaks because of any change in the upstream system, AI comes back into play and fixes the automation itself. This is really the pattern that we are seeing emerge in an enterprise. AI is actually creating the software that runs an enterprise.
This software cannot behave incorrectly because it cannot change its behavior in real time. It's not a probabilistic technology. Even if the software is created by AI, we can audit it, and we can have humans read it and validate it. I can have tests that, for a certain input, will guarantee that the software will behave in the same way. That, again, makes this pattern extremely powerful. You use AI to create software that runs the enterprise in a predictable, governed, auditable way.
How many engineers do you have today?
Maybe more than 1,000.
More than 1,000 engineers?
Yeah.
How many people do you have?
Around 4,000.
You have 4,000 people?
Yes. Why so?
I find 18 a fucking nightmare. And what? 4,000. Oh, my God—no, do you have too many?
4. AI Transformation Needs A Work Map
It's a complicated question because I think I would answer it with what is in my book. The more we transform our companies using AI, the more we have to transform our workforce at the same time.
I've never hidden from my employees that there will be a transformation in the company. But I told them upfront, "Guys, we are not doing anything stupid. We are not just using AI as a pretext to cut part of the company. We need to do the transformation at the same time as we successfully adopt AI in an enterprise."
That's actually another point that I discovered while writing this book. I was looking deeply at jobs, what jobs can be affected by AI, what jobs can be enhanced, and what this transformation is going to look like. One of the things that seems very simple in retrospect is that people don't have simple jobs that can be defined on a sheet of paper.
Every job has some kind of measurable outcome, and this is the outcome that people are hired for. But there is another outcome that is part of the institutional strength. Think about my deep relationship with the customer. It's not necessarily part of the numbers that I'm producing, but it's maybe what makes this customer stick to my technology. This is a different outcome.
If I'm going blindly and cut on a number of As and say, because AI is going to replace them, AI can call the customers and write emails, I don't think AI can supplement the human connections and the trust. So this is a different outcome of the job, and it reflects on every employee and every type of role in the company.
To me, an enterprise should have a ledger where they actually understand what people are doing besides their main definition of the role. Only after they have this ledger and understanding can they look at what the AI transformation looks like. What kind of jobs will be affected? How can I move people from one job to another because they still carry some kind of cultural aspect of the enterprise?
I don't think it's a simple problem where AI is going to cut 20% of the company. Let's do a RIF, take 20% out, and then increase AI adoption. On the contrary, when you do this blindly, you risk hollowing out the enterprise of exactly the same talent that thrived during AI.
Let me give you an example. In the book, I call this the "credentialed middle." That was the type of people who were prevalent in every enterprise. If you think of our education system and the way we hire people, we hire based on deep expertise in a particular domain, credentialed expertise. This is exactly the type of expertise that might not be needed as much because AI can really help.
You will need fewer of these experts, but you will need more people who have initiative, who are capable of maintaining a relationship with the customer, who can be mentors for new employees, and who bear the cultural aspect of the enterprise. It's counterintuitive because you will tend to cut those people who are not the biggest experts in the domain, but you will cut exactly what you will need to bring the AI to supplement these experts.
But maybe you just need fewer of them. I was speaking to a lawyer today, and I said, "How big is your trainee program?" He goes, "Well, historically it was 25." I said, "Wow, that's a lot of trainees." And he said, "Yeah, but this year it'll be 4."
I 100% agree. We will need fewer people in probably most of the roles, but the main question is which ones? How do you choose?
I think you can choose quite simply based on where there is verifiability. Finance and accounting are quite clear about what is right and what is wrong.
I disagree with you. It's not about verifiability; it's about whether the work has been defined in a frame set by other people. If the frame is clear, then AI can understand the frame.
But the frame is clear. Finance and accounting—they do your expenses.
No, it's not. This is one of the domains where the frame is not clear. When I receive an invoice or an order from a customer, I can treat it differently. There aren't always rules there.
I know that for this customer, NVIDIA is going to ship with priority to OpenAI. Maybe they have a rule that says, "Yes, my first chips go there." Maybe they don't. If that rule isn't captured in a frame, AI cannot learn it.
This is why you need to create this manual. We call this manual the map of work. You need to hand the map of work to AI in order to be successful.
What do you mean by the map of work?
It basically captures how the work works, how the work happens in an enterprise. This is the map of work. It's all the workflows, all the exceptions, all the procedures, and all the systems that you use in order to fulfill the goal of a process.
Sure, but then you have a head of finance who sits on top of 30 agents, and exactly when NVIDIA comes back and says, "Whoa, whoa, whoa. We're your biggest buyer, and we have special terms on our payments," they go, "Yeah, sure. That's right. Don't worry about it."
But you come to my point. Even in finance, you cannot replace everybody, so—
Not everyone, but you've got 1 person or 2 people.
You'll need a certain number of people. The thing is, if you have X number of people, how do you understand which of them stay and which of them have to go to do different jobs? How do you know?
Ideally, you will get the people who have AI literacy and can display initiative. AI cannot exhibit initiative in the human sense. So out of this number of people that you want to keep, you want to keep the people who display the most initiative.
Even a finance person treating an invoice for a customer contributes to the culture and how my enterprise is regarded. How can I distinguish between that person and another person who cares less about how they treat the customers? The second person is more prone to being displaced by AI.
This is the ledger that I think enterprises have to create in order to understand the different outputs of people. They need to judge people by this hard-to-define output.
So how does one do that, then? For the illegible data that isn't captured within companies, how do we do that? Zuck and Facebook have talked about monitoring every single action that's on the screens of employees. I don't think that captures the tone of a call, the warm text afterward to a customer, or the invisible data. How do we think about capturing the data that shows value but that we don't capture?
I think this is the crux of the problem. And this is—
And here, you're a true philosopher.
Yeah, and this is where we put a lot of effort as a company. We are introducing a new technology that we call cartography, and cartography is a discipline. It's a discipline to help companies surface all the information about how the work is done and help them create this map of work.
One big, important part of cartography is to investigate what people are doing on their desktops. We have a product that we call the Cartographer Agent that can interview real subject-matter experts, have them record what they are doing, and interview them in real time.
If you interview a finance person, it can ask, "Why did you change this invoice when the ZIP code was different? Why did you choose a different path? Tell me." They can start surfacing all of these exceptions. So that's real agents interviewing real people. It's pretty cool stuff.
Then you consolidate data from multiple people, and we create the process maps. We show them how the work works in real time. After this map of work that shows the work as is, you can use our coding agents and come up with an idea of how you should transform the process. You transform the process by printing software.
You start with a process, and at point A, it's fully manual. Of course, you have enterprise systems like systems of record, but people operate the systems. I think the goal of any enterprise is to have fewer people operating the systems and more automation and agentic AI operating the systems.
Do you not fear pushback from people working in the company who are aware that you are watching what they do to replace them? That is what Zuck got.
This is inevitable. It very much depends on how you pitch the company. In UiPath, I think it was important to tell people, again, "Guys, we are not doing anything stupid. We are not doing any mass extinction under the pretense of AI. But transformation is inevitable, and you guys have to transform, and everybody will get a chance. The people who become more literate in AI will have a better chance not only here, but in the future, in any other job."
That's the message that I think everybody should get.
Did they respond to it? Did you see AI adoption go through the roof after that?
I think the response is good. AI adoption requires more cycles than just discovering the process and getting people's input on this. But look, at the beginning of the year, the fear of people across the industry—not only in my company, but in many companies—was off the charts: the fear of being completely replaced.
Now, I think people are starting to get a better understanding of the durability of their jobs and of this AI diffusion in enterprises, which can happen at a slower pace and one process at a time, because these millions of Einsteins are not hireable yet.
One of the companies we invest in, Macaw, an AI data provider—Brandon Phoody, the CEO, tweeted yesterday that they spend 3× the spend of human salaries on inference. What percentage or multiple would you say you spend on inference relative to human salaries?
I personally don't care about it. And let me tell you something.
It’s a simple hypothesis. If work at the quality of a human can be done by a machine, I will hire a machine today, even if it’s more expensive than a human. Human costs will only increase, and humans bring errors into the picture, while the cost of machines will decrease. So I will have a competitive advantage compared to people who stick to humans. I think everyone will do this.
I don’t think the cost of tokens will be the real question if you replace a person with AI. But the real problem today is that AI cannot replace a person because, again, if you bring me an Einstein who can replace me, I will happily go on any vacation in the world, but this Einstein doesn’t yet exist. I would like to be interviewed and have this podcast with another Einstein. This thing doesn’t exist today. That’s the reality, so let’s call it a reality. Let’s call a spade a spade.
Maybe this technology will emerge and, somehow, Einsteins that embody a person, that have will, that get transformed on the job, learn on the job, and have the capability of reasoning and imagination will exist. Of course, all the jobs will go extinct.
I have a show with Jason Lemkin from SaaStr. He’s cut his team from 25 to 2. If he were in my seat now, he would say, “No, no, no, it does. I replaced my VP of finance. I replaced my VP of marketing, and actually, the AI is better.”
I want to see this man. I’ve heard of companies that replaced hundreds of support people in the past, and now they are rehiring these people. I think until this model is proven at scale—not in a particular industry for a particular guy—I don’t think we can extrapolate from 1 data point that it’s going to go across industries.
You talked about extrapolation and over-exaggeration. The SaaSpocalypse was very real. We’re going to vibe-code everything. Did you vibe-code tools out?
5. Vibe Coding Meets Production
Look, it was amazing. Yes, we did, but it wasn’t an extraordinary success. Initially, it seemed extraordinary, but when we tried to put it into production, we started to see some real bottlenecks with these tools. You need to have a lot of things to maintain: connectors, permissions, audit, security.
Taking software from a prototype to production is actually where the work is. It’s not necessarily the writing of code. Writing code is fun, but that’s not where you can really make the difference. I think it’s much easier today to make a prototype. A prototype is so easy, but then you have to iterate to make the prototype work in production, and this is the testing and everything else.
We were trying to replace a procurement tool by writing it ourselves, and I think we had a lot of success initially. It was written only by AI. But then, when it comes to these self-improvement loops, we don’t trust it—we don’t have enough tests, and we don’t have enough trust to put this tool into production 100%. In our experience, humans have to intervene a lot in how this vibe-coded tool works.
For instance, the database schema that the vibe-coded tool created was completely bogus. A human has to come and create the structure. So right now, you are not at the point where you will have a business user who understands a problem and will vibe-code a tool. You will still need to bring in engineers and people to maintain it. So it’s a long process. Eventually, you will end up paying probably as much as, if not more than, the tool you replace, while you keep some of your good and best people’s bandwidth occupied.
Would you buy Salesforce today?
6. Public Markets Meet AI Infrastructure
I would buy Salesforce as a system of record.
As a stock?
As a stock. Look, I invest in software as a category, and I think I made a good investment a few months ago because I bought at the bottom of the SaaSpocalypse. Even our own stock has been doing better. But the markets today are driven so much by sentiment and not by value. So it’s kind of hard for me to make a judgment of an individual company. But I don’t think Salesforce can be replaced by vibe coding, if that’s the question.
I don’t understand why a company would go public today. If you think about the 2 drivers of being public, number 1 is liquidity for employees and shareholders. Stripe and many companies are able to have liquid stock in private markets. Number 2 is the ability to have M&A—a tradable asset that you can buy with. I mean, many private companies are able to buy other companies with private stock. Stripe was going to do PayPal with private stock for $60 billion. So that’s not a barrier.
The casinoization of public markets, as you said, with current stock markets being sentiment-driven—I don’t understand why one would.
But Harry, let’s not make a confusion between some very exceptional companies and most companies that are out there. There are so many companies that are zombies right now. These 2021 zombies would fare better in the public market right now. At least their investors will have a way of exiting, and their employees will have a way to make some money. Nowadays, all of them are sitting on paper, okay? But public markets will confront them with the reality of their real valuation. Why is Anthropic trying to do an IPO in the end?
Well, but they’re unique companies, alongside OpenAI, which just has to go public because they’ve exhausted all the private funding that exists. They’re extraordinary companies because they just need too much money.
Do you believe all the investors in OpenAI and Anthropic will stay in the companies for years to come?
No. I think some will, but some—
Some will, of course, but I think we will see an exodus. Honestly, I don’t believe in a $2 trillion valuation or whatever. Maybe they will reach $5 trillion. Because if I buy at $2 trillion, I need to have a path to $5 trillion.
But if they went public at $2 trillion, would you sell?
Anthropic in particular, I wouldn’t. Do you remember on our last podcast, I think you asked me which company I bet on, and I said Anthropic? Anthropic was worth a $60 billion market cap. Maybe I was stupid; I didn’t invest.
You would have made more money on that than—
Yeah.
Salesforce or ServiceNow or whatever.
Yeah, 100%. So I would not sell, and I think it’s the same with OpenAI. I think OpenAI has caught up quite nicely, and I use them interchangeably right now.
I just think we’re in a market where the big get bigger and value concentrates more than ever.
But the real question is, Harry, would I buy at $2 trillion? That’s my real question. Right now, I need to see their real numbers to understand if I will put money in their IPO.
I’m going to get in so much trouble for this. I think AI is quite like Bitcoin, in just the way that it’s very difficult to determine what application is going to win, what wallet is going to win, and what usage is going to win. If that is the case, buy the underlying infrastructure that you know is going to be there.
For me, I agree with you. I don’t know if Claude is going to be better than the next Codex. I don’t know if Cursor is going to come out with something fucking amazing. But I do know that Jensen is going to be sitting there going, “Here’s another chip. Here’s another chip. Here’s another…” Great.
Yeah, but I think Jensen is bound by the success of open source. If Anthropic and OpenAI become a duopoly—and I think their TAM is in the trillions; it’s basically the work—they will print their own chips, man. Honestly, it’s not such a big deal, in the end, to print chips.
Of course. I mean, OpenAI are doing Jalapeño—
Yeah.
—and Anthropic are doing their own chips.
Exactly.
Yeah.
So I don’t think Jensen will be doing so well if they have the single biggest providers, and the source of truth and light of God will come only from Anthropic and OpenAI. Therefore, open source should succeed.
I absolutely agree, which is why I think Jensen is doing the open letter, which everyone signed, encouraging open source.
Absolutely.
Hugging Face.
Hugging Face. Yes.
Why?
Because I think it encourages open source. It hosts all the open-source models. It’s putting the money where the money is for his company.
Totally. It also makes a neutral provider no longer neutral. Bias. Obviously, they have Nemotron, and they have their own models now as well. You could say there’s a loss of independence now that it’s owned by NVIDIA.
I think Nemotron is still a small cog in the picture. I think it’s valuable, but it’s not at the same chip size as the others.
Do you worry about the round-tripping revenue? Everyone talks about NVIDIA investing here, buying here, and the circular economy that comes from Oracle and OpenAI. Do you think that’s overblown?
It can be, because every major infrastructure in history has been overbuilt. I think there’s a simple explanation. I was thinking, why is every infrastructure overbuilt? Because you have to make sure you get the biggest piece of the opportunity. If the opportunity is big, it doesn’t matter. You build a little bit more than is necessary.
So it’s clear now that there is only 100% of the pie, and people are building 200% of the pie. There will be losers.
Do you not think, though, this is the first innovation where we are significantly underbuilt? If you look at the constraints now, you’re right. In prior technology cycles, we overbuilt the supply side and the demand side was lagging behind.
Now energy is a massive constraint. We have water, data centers, regulation, and policy. We have a significant hindrance to the supply side, and we are underbuilt, not overbuilt, which is why every ounce of compute is taken.
Yes, but are we underbuilt to the extent of the trillions coming into the infrastructure? I don’t know the answer to this. Everything happens on the premise that AI is going to replace human work on a really large scale.
We need to see the timing of this replacement and transformation. It’s a big difference if it’s coming in 10 years versus the next 2 years. I don’t think it’s overbuilt for the next decade, but it might be overbuilt for the next 3 years. Stock markets and capital can be merciless.
Maybe I’m a childish optimist, but I saw Andrej Karpathy say that he used coding tools for 20% of the work, and then 6 months later he said that they did 80% of the work and he helped with 20%.
We’re investors in Lawgora. I interviewed lawyers when we did that deal, and they all said to me, “You’re such tech bros. You think you can replace us. Ha ha, we went to law school.”
I interviewed them 2 weeks ago. Every single one of the 15 said they would be severely unhappy if it were taken away, with most of them saying they hadn’t written a document in 6 months.
Harry, you should read my book, my friend. It answers—
I did.
Exactly the same questions.
I loved it.
When the frame that a person operates in is really well defined by someone else, like in law, AI can be devastating in its impact. When the frame is not as clear and there are so many exceptions that are custom-made for an enterprise—
Dude, you’re about to spend 2 days with a lawyer who is my girlfriend. She will tell you that law is highly ambiguous and subjective in terms of writing styles.
So is human language, and AI understands it perfectly. It understands every freaking nuance of human sensitivities. As long as it’s documented and well-defined, there is a manual for the freaking law, and AI is amazing.
When there is no manual, AI is not amazing, and it doesn’t work. That’s the huge difference.
The legal industry in the US is $300 billion. It’s a lot. If you think about how much labor could be replaced by that, I think 30% would be reasonable. That would be $90 billion of available revenue.
Yes, but that’s not going to convert into token revenue. Maybe out of $90 billion, companies might charge 10%. Maybe it’s a $10 billion total opportunity in tokens.
Am I mistaken to invest in Lagora, and are the Harvey investors mistaken to invest in Harvey if it’s $10 billion, not $90 billion?
7. Workflows Capture The Real Value
What I can tell you is that, from a law perspective, open-source models and frontier models will do just fine. Maybe they will also do the custom workflows around the legal process, which is really valuable.
To me, that’s also a big part of my thesis. Models are interchangeable, but the workflow—the map of work and the workflows around the map of work—is where the real value is.
If Harvey and Lagora are doing this, they really map the work, create the workflow, and create a legal department for me. Of course, it’s a much bigger value that they capture. But if it’s only to get a legal opinion—a call to a model—that’s not going to be a $100 billion market. 100%.
What percentage of token traffic do you think will go through open versus closed models in 12 months?
To me, I think the question is different: What percentage of the traffic will go to truly frontier models like Astra or Fable versus very cost-efficient models? For enterprise work, my prediction is that 90% of the flow will go to very cost-efficient models.
I don’t think you need frontier-level quality models for most operational work.
Just to be clear, then, we will actually still use the core providers, which are OpenAI and Anthropic. It’ll just be deprecated older models.
I will still use Anthropic and OpenAI with their cost-efficient models, but I will have a verifiable backup in open source all the time. As a responsible enterprise, I should be able to switch models. I cannot be locked in.
I’m just checking my portfolio against your brain. I believe strongly in open models, and I think that every company—not every company, but mid- to large-scale companies—will have their own model, own their own intelligence, and feed their own data into it.
I think that goes to the statement of owning your own intelligence, not renting it. That’s why we invested in Fireworks, and I believe in the open model ecosystem.
Yes. I’m a big fan of Fireworks, and we are using them quite a bit.
Do you like them?
Yes, we like them a lot. I’m a big believer that an enterprise should distribute its bets, and one of the bets should be on open source and, very importantly, on this map of work.
Think about it: If I want to train my own model with my own “who am I,” I need to have this “who am I” very well documented. I need to create this manual, because I’m training one model today. But in the next 2 months, there will be another, better base model coming into the picture.
How can I do transfer learning from my old model into the new model if I don’t have the data and the exact manual? I cannot. Otherwise, there are terrible losses when I do this.
The real investment for an enterprise is creating this map of work that documents how they actually work. With this, they can train their own models, whether it’s in Fireworks or another provider; it doesn’t matter. But this is their IP, and this is their core data. Make sense?
So you do believe that companies—and a lot of them—will have their own models with their own data?
I do believe that they will at least have their own models as a backup to frontier models. To me, where I’m not clear is whether I can provide the same cost efficiency with my own model versus a cost-efficient model from Anthropic and OpenAI.
I think these guys are in a position to truly optimize large infrastructure. Part of their business model will be to deliver more intelligence per dollar than even I can squeeze from my own models.
If you have highly specific data that is exact to the request you have, which is your data, I think you’ll get more token efficiency with your own model than you would with an optimized frontier model.
Only if you are training your models might that be true, and only if Fireworks can deliver at a large scale and in a very optimized way.
Would you invest in Fireworks at $15 billion?
Probably, yes. If this hypothesis of open models is true, which I believe it is, I think they are undervalued.
I think they will have to get into this big game of securing compute very soon. Because if they don’t secure compute, I don’t understand how they can give me inference at the scale that I want. What do you think? You invested in them.
I did. I think you’re absolutely right that they need to move into the compute layer, and I think Lynn is doing that, I’m sure, very soon.
So they will have to raise tens of billions now.
And I will be there.
Yes.
No, and she did that at Facebook. I think that’s unique to this team. They secured compute.
They’re a great team. We really like them.
It’s a great fucking team.
And we worked with them before the big hype.
I also think the data providers are massively underpriced and underappreciated—Macaw and Surge in particular. Everyone’s like, “Oh, they’re commodities. You’re just buying data.” Data is the most important thing for model quality.
But what’s the difference? I think one thing is storage, and one thing is understanding of the data. If I have storage, I can have a tape and put data on the tape. Would you invest in a tape company? I don’t think so.
You need to invest in the intelligence that understands the data, feeds the model, and extracts the right data at the right time, feeding the model with the data that is needed, with the context.
Because if you have just data but don’t have a way to create really good context to give the model when it asks something, it’s useless.
Well, I think you would say that they have more data than anyone else across more categories than anyone else. When the model requests highly specific data, because of the breadth of their library, they’re able to provide it in a way that others aren’t.
If it’s their own data and it’s valuable for models, I’m sure the models will buy the data in an instant.
Can I ask you, what have you changed your mind on most in the last 12 months?
I didn’t understand the necessity to have a manual in order to work. That was maybe the biggest breakthrough in my understanding: Every time I’m running a query toward AI, AI should have at its disposal the entire way my company works, or this particular process works.
When I realized this, I also understood that this is the biggest differentiation between memory and true learning. This is how we started the discussion, and I'm not sure I really made a point, but there's a huge difference between just laying something down, having a scratch pad, and being transformed by an experience. That's the thing that I realized the most, and I think I also realized what is kind of human for us, because I experienced a lot with AI writing—not code, but writing a book.
I've been through different styles, and I understood a lot about how to prompt them. AI doesn't have a style, and you realize why they don't have a style: they're an averager of anything. In order to have a style, you need to have a body. You need to have individuality, because we are the choices that we make and the choices that we don't make, in a sense.
You need to be transformed, because otherwise I can just ask AI, “Read this book and write in the spirit of this author,” and it's not really working, because you need to be transformed by the experience. To me, this is going to be the biggest breakthrough in AI technology: when I can have models the size of Mythos being transformed on the job, being put in a laptop. It might be possible. Who knows? You know the pace of technology.
Maybe 20 years from now, I can have a 10-trillion-parameter model that is my own model and is getting transformed along with me. But we need to see. I think there might be a few series of innovations to get there, because I want to give you also an interesting data point.
AI is solving very interesting math problems that humans hadn't solved before, right now. But AI still isn't capable of creating frameworks. Relativity is a framework. I was thinking, why is that?
I think one of the main reasons is related to this not being transformed when you're on the job. When I'm writing a book, I'm being transformed by the act of writing this book. Every time I'm writing something down, there's something in me that changes that isn't necessarily the memory of a thing. It's me who is changing.
Einstein was changed by his experience thinking about the speed of light, about what happens when you go behind the light. It's not like Einstein wrote it down and then, every time he thought again, he rewrote a piece of paper. No, he gradually became a different Einstein from the one who started thinking about the problem, when he created this framework of relativity.
Models don't work this way. Even if I put in a swarm of agents and everything, they have to write down everything. They aren't being transformed by the process. Therefore, in the end, it's very difficult. They will have this context, but it's very hard to go beyond a 1-million-token context window.
A framework might require a transformation as you work on that framework. It's a different way of learning from pure memory. That's the argument I want to make as clearly as possible.
Are you optimistic for your children?
I'm extremely optimistic for myself, Larry. Therefore, I'm optimistic for my children. I don't want to sound like an AI doomer, because I believe that—
I don't think you do.
I am.
No, I don't think you do. I sound like a doomer in a way. I think we'll have a lot more job loss. I think it will happen a lot quicker. I think we're seeing it in real time.
I'm much more optimistic that we won't have so much, because based on my own experience with AI, I don't think the diffusion is as fast as you imagine, particularly because enterprises have to document their processes in much greater detail.
Can I just ask? We're both Europeans, and we're both sitting in London. I don't know how to say this, but we don't matter anymore. Just being blunt. Do you think that gets better or worse in the next 3 to 5 years?
8. Europe Needs Sovereign AI
Yes, man. It's hard to admit the reality, but from a technology standpoint, I think we're largely irrelevant. But it's so stupid, because the biggest producer of machines that make chips is based in Europe.
It's ASML.
Yeah. We could have made these chips in Europe. Some of the most brilliant minds who build AI—even if you think of Dario, Sam, all of them are of European origin. Ilya. We have the talent. We have the technology to build the machines, but somehow we're losing it, and it's very stupid.
Do you see a difference in work ethic having a team in the US and the UK?
Yes. I have experience with teams in the UK, and at 5 PM, they're all in the pub.
Why is that? Because money matters more?
I think culture matters more than money. It's a more dynamic culture. I don't think I would have succeeded in Europe the way I did in the US. I'm a European, but as an entrepreneur, I'm American. This is what I tell everybody, so my formation is in the American school of entrepreneurship, even if I started my company here.
Listen, I get it. You look at Lagora, you look at ElevenLabs, and some of the best companies to come out of Europe in the last few years. If you think the revenue machine is anywhere but in America, you're lying to yourself.
It's an easier-to-access revenue machine in America than in Europe, clearly.
Faster. It's easier. The teams have scaled go-to-market functions before. I completely agree with you.
American companies are making larger bets on vision without waiting for so many proof points as European companies. Even people in middle management can make sizable million-dollar bets on new technologies in the US. I haven't seen this appetite in Europe.
I don't want to ask this, but I'm interested. If you were to advise a young European entrepreneur today, would you say to go to the US?
Yes. That's the sad reality. Unless they build for a specific market with some specificity in mind, if they build a universal technology, they'll have a better chance to succeed in the US. There will be many successful European companies coming out of this. Maybe not as frontier labs, but I think for the application of AI—
Do you buy sovereignty as an argument?
Yes. I think it's an important one.
Energy sovereignty, model sovereignty.
Yes, 100%. All European customers right now would prefer on-premises software, model sovereignty, and model optionality. This is a big business that's coming here.
Look, I talk to our friends at Fireworks, and I actually try to convince them to make their software available on-premises. Right now, they're in “show me the money” mode, but I can tell them, “Guys, this is a big business. You need to prove first, because this is Europe.” Show them the technology, and the money will come.
How much revenue does UiPath do today?
I think it's public data. We're at $1.6 billion, growing last year by 14%.
Jason Lemkin taught me that unless you're growing 20% or more, you're just fucked in the public market. It's grow or die, and it's a horrible reality. I'm not condoning it. It's horrible.
Yeah.
Is that right?
Yeah, because I think the public markets are very confused right now about who the AI winners or losers are. If you don't show serious growth and traction, they automatically put you into the AI losers category without looking deeply into the business. There are so many hundreds of software companies in the public market, so it's hard to look at each of them.
If I were to flip it on you, we'll do a final one for quickfire: what is the bull case for UiPath being a $50 billion company?
Think about it: Gartner released its new BOOT Magic Quadrant—Business Orchestration and Automation Technologies. We are one of the leaders. We moved from a challenger to a leader in the last year. It shows that, as a company, we made this transition from an RPA and automation technology into an orchestration and automation technology.
There are all the arguments in the world that this is really required in order to create this new enterprise that is AI-powered. This idea that you can have an AI agent that runs everything for you from top-level processes, orchestrates and automates everything by magic—I think it's something that people have stopped believing. You need to have an underpinning orchestration and automation technology, and this map of work that I talked about, in order to power your processes.
This is what we have. It's not only me saying it; it's Gartner, Forrester, and industry analysts who are very bullish on us. So that's really the argument right now.
This asymmetry that AI is creating right now is more obvious: printing software that runs your processes has become much easier than it was a year ago. Creating an AI agent that runs your software is as difficult as it was a year ago.
So you make a huge investment in building this software, capturing the enterprise context that we call the map of work, and putting this enterprise context inside these rails that I named orchestration and automation—the map and rails. The map is the context; the rails are the orchestration and automation.
You put them in the same platform, and then you can assign an agent to do work. You tell the agent, “This is the reality. These are the rails you can use. This is the map that describes how to use these rails. This is the goal.”
That's the way you can have control on top, and your agents cannot go rogue. No sane enterprise right now will put in a swarm of agents and just ask them, “Do my financial accounting for me.” Who knows? Maybe they'll attack your competitor—
Rogue AI. I completely get you. What is the bear case?
I think the bear case is that AI will somehow become a genius. Token costs will be next to zero. We’ll have literally millions of Einsteins in a data center—but Einsteins in a true sense, not only in reasoning, but in the sense of replacing a person. I can assign them to every task in an enterprise, and they will just do it. That’s the bear case against us.
I mean, token costs have gone from $60 to $1 per million tokens. So token costs will go to nothing.
It’s possible.
Yeah.
This is why I told you I would not stop an investment right now based on token cost.
Dude, I could talk to you all day. I’d love to do a quick-fire with you. I’ll say a short statement, and you give me your immediate thoughts, okay?
Mm-hmm.
What’s the hardest thing about your job today as CEO of UiPath?
It’s aligning people. There are so many different personalities, and pride and ego come into play. This is the hardest.
What has changed most about how you work as a CEO because of AI?
I’m spending maybe half of my day right now alone with myself in Visual Studio Code, working with Claude and ChatGPT. I have way more leverage on my company than before because we completely changed the way we operate.
Most people, when they came to me with an idea a year ago, would come with a deck, and it was very hard even to prepare this deck. Now everyone is going to come with a Markdown file, and I can put it into a giant strategy folder where I have AI agents working with it. I put this document in my folder, and then I can ask intelligent questions.
If you had unlimited resources and zero retribution from Wall Street, what would you do that you’re not doing?
Maybe I would try to build my own frontier model.
NVIDIA in 3 years’ time: will it be above $7.5 trillion? It’s at—
We’re there today. Five.
Five-six. $5.6 trillion, yeah.
I can easily imagine a 40% run for NVIDIA. I would bet more on NVIDIA rather than Anthropic being a $7 billion company. Really, of course. Billions are nothing today.
Billions are nothing today.
Yes.
You said something on a show that we did before, and it was one of the most resonant things I’ve ever heard on a show. You said, “I think a lot of people think they want to be me, but sometimes it’s quite lonely alone in my head,” and I always remember this because I often feel the same. What would you advise founders who feel lonely in their heads and struggle with that today?
I think they should surround themselves with their best friends from childhood, perhaps, and have more frequent chats with them because they are the people who can relate most to who they were before, and they can see them as part of the transformation. It’s a nice thing to do anyway.
You’ll still be lonely, but you’ll have a sense of some kind of continuity in your life. I find one of the most rewarding parts of my life is chatting and being with friends and family. This is really where you get a lot of relief from loneliness.
A final one. What are you most excited about when you look ahead? My mother’s got MS, and I’m really excited about some of the breakthroughs we’ll see with chronic conditions and their treatment.
Yeah. I’m very excited about longevity. You know, my friend, I’m almost twice your age.
You’re not quite.
I don’t do any kind of gym or anything.
Dude, are you kidding me?
And—
No. What are you doing longevity-wise?
I’m doing quite a lot. I got into peptides and supplements. Man, I think I’m taking around 60 different supplements a day and 3 or 4 peptides.
60 supplements?
Seriously, yeah.
60?
60, yeah.
What the fuck are you taking?
All of them have been recommended and vetted by AI.
What? That’s extraordinary. I mean, you look incredibly young, but 60 supplements. Are they in pills? Because I do the Longevity Shape from Bryan Johnson, which is 60-in-1, and I just have it every morning. You actually have 60 separate supplements?
I have a lot of pills. I also do some in the form of powders, but yeah, I have— I’m going to show you tomorrow all of them. I have little bags for the day, like AM 1, AM 2, AM 3.
That’s extraordinary.
I know.
That’s extraordinary.
It’s very dorky of me. Yes.
Do peptides make you feel better?
I think they’re supposed to make me feel better in the long term, but honestly, I feel way better even than 10 years ago. Reducing booze quite a lot helped, and I know you’re a big fan of booze.
You don’t still drink, do you? Do you still drink?
I drink a lot less these days.
I love that. Dude, this has been so much fun. Thank you so much for putting up with my meandering. When does the book come out?
It’s already available for download. I’m also printing a few copies. We have our big Fusion event coming in a couple of weeks, and I’m distributing a copy to everybody coming.
Dude, this has been a pleasure. I’m going to get a copy. I’m going to get a physical copy because I’m old too—
Yeah.
—and so I like reading.
It’s my gift to you, Harry, of course.
There we go, dude. Thank you so much.
Thank you, man.