Dr. Niklas: Venture Grade logoDr. Niklas: Venture Grade
November 10, 2026 · 42 min · 31 min read
DN Briefing #2: A Brain Needs a Body, The 2% Market & A $1T Anthropic (w/ Martin Tobias) cover

DN Briefing #2: A Brain Needs a Body, The 2% Market & A $1T Anthropic (w/ Martin Tobias)

With Martin Tobias · hosted by Dr. Niklas

"Claude Code is one harness, one market, one use case for this brain. There are 1.5 billion knowledge workers. Two percent of them are programmers."

The DN Briefing: Martin Tobias (Incisive Ventures) and I take the biggest stories in startups, tech, and venture and ask what they actually mean for founders and early-stage investors.

This time: why harnesses are a different business from wrappers, why robots are just harnesses with bodies, the hair salons and railroad yards getting their first real software, Anthropic's valuation cut and why a $1T IPO might be the better outcome, the neo-cloud debt problem, open-weight models at 10% of the cost, and the one list that tells you where not to build.

In this episode:

• Brain and Body: A wrapper prompts the model. A harness gives it eyes, hands, integrations, and a workflow.

• Muscle Work, Then Cognitive Work: What happened to physical labor 150 years ago is happening to knowledge work now.

• Robots Are Harnesses Too: Why delivery bots and warehouse robots scale before the Jetsons maid.

• New Software for Old Industries: Aura300 for hair salons, Quintess for railroad maintenance, and the reluctant technology buyer.

• The Descriptive World: ChatGPT trained everyone to type what they want into a box. B2B is next.

• Neo Clouds and Debt: Why Martin shorted CoreWeave, and why balance sheets beat leverage for building infrastructure.

• The Case for a $1T Anthropic: 98% of Google's value was created after the IPO. Public buyers need upside too.

• The Fiber Lesson: Dot-com investors lost on fiber, and the whole internet got built on it.

• The 2% Market: Anthropic's growth runs on one harness and one job. The other 98% of knowledge work is open.

• Open Weights at 10%: OpenRouter's majority of tokens now go to open models, and frontier revenue still climbs.

• 72 API Calls, 90% Second Tier: How Martin routes his own product, and why every app will.

• The 10 Jobs List: Anthropic published its roadmap. Build the vertical harness outside it.

• The Founder Martin Wants to Meet: Ten years in an industry, bought every tool, had to build it themselves.

🎧 Full episode on all podcast platforms

💬 Which old industry gets its harness first: hair salons, railroads, or something nobody in Silicon Valley has heard of? Let us know in the comments!

🔔 Please like and subscribe! Every subscriber helps our channel grow.

#DN #DNBriefing #Harnesses #AIAgents #MartinTobias #IncisiveVentures #Anthropic #VerticalAI #OpenWeights #VentureCapital

Timestamps:

0:00 Claude Code is 2% of the market

0:19 Intro

0:31 Harnesses vs wrappers: giving the LLM a body

3:58 Model as CPU, harness as OS, and the robotics question

5:02 Muscle work moved to machines, cognitive work is next

9:56 Waymo is a robot, and where the software opportunity is

10:38 New software for old industries: hair salons and railroads

15:44 The descriptive world: ChatGPT trained everyone to type in a box

17:28 From Q&A bot to work product, voice, and robot interfaces

19:50 Anthropic's valuation cut and the neo-cloud question

20:16 Debt vs balance sheets: why Martin shorted CoreWeave, and the case for a $1T IPO

26:09 One trillion is still an outstanding result

27:54 The dot-com fiber lesson

29:12 One harness, one market: Claude Code is 2% of knowledge work

31:41 Compute is short: TPUs cost more than a year ago

32:06 Open-weight models at 10% of the cost

33:26 OpenRouter flips to open models; 72 API calls, 90% second tier

35:42 Enough value at the frontier: science and programming

36:35 Anthropic's 10 jobs list is their product roadmap

37:00 Build the vertical harness: Harvey, Cursor, and 28 POS systems

40:28 The founder Martin wants to meet

Transcript42 turns

Martin Tobias - Incisive Ventu:But if you figure like 70-80% of their revenue is from Claude Code, which is basically one harness, one market, one use case for this brain. and that one use case, there's 1.5 billion knowledge workers. There's only two percent of those knowledge workers who are programmers.

Niklas:Hi, and a huge welcome to you, my lovely listener. So glad you're here. Today you're joining me for another briefing with Martin. Martin, so lovely to have you.

Martin Tobias - Incisive Ventu:Hi, nice to see you again.

Niklas:B to B harnesses versus LLM wrappers. I think you've been quite vocal about not investing into like thin wrappers in the beginning. Now we are seeing harnesses that are really capable coming up and it's a totally different story. Maybe give me your take on it. What do you think?

Martin Tobias - Incisive Ventu:Well I think and and I and I posted a little bit about this, The first thing we should do is level set about what the difference between in my mind between a wrapper and a harness is. In my mind, a wrapper is, you know, in the beginning of a lot of this LLM stuff, it was like a better way to prompt the model. or something that had had it was mostly prompt engineering is what rappers were, like saying you're this and that and the other thing, and here's how the output should look. So it was prompting the Model of it in in a specific way to get a specific output. That's what most rappers were. But what harnesses are is not just a prompt, but if you think of the LLM as a brain, a harness is giving the LLM a body. So you've got inner integrations to other data sources, you've got eyes and ears and arms, you've got a browser that the LLM can use to do something, you've got a defined workflow where it understands how to do something one two three you you've got hands to to complete things so harnesses have a lot more integrations and are like a body to a mind to actually complete things and in that case you have a lot more value in defining a harness that can do something for example Claude code is a harness against the LLM The and and and what it what does that harness have that the LM doesn't have? It has an integration to GitHub to find your source of truth. It has you know been trained on specific coding things like how to debug what the difference between a senior you know database engineer and a UI engineer is, you know, how to run a test and and set up specific unit tests and user tests and things like that. So anyway, so it understands a bunch of Of things about how to be a programmer and what tools a programmer needs, which are in addition to the intelligence that the brain needs to run those different processes. And so that's the main difference between a wrapper and a harness. And I would argue that Claude Code is the first harness that really has gotten lots of adoption and it solves a specific workflow. And if you think about it, there are harnesses to get specific jobs done in in tens of thousands of industries that require specific integrations and knowledge of specific workflows and specific sub-uh tasks to complete those vertical workflows. And since you're getting work that's actually done, instead of just adding intelligence to a work, adding a brain, you're adding a body. That's why I'm really excited about harnesses versus rapids. Yeah.

Niklas:Yeah, for for me, I think if you look at it, for me the LLM becomes the CPU. So it's the processor. It's just a different one than the one that we know. So the the harness, if you look at it, it's some kind of operating system. So this is where everything else plugs in. So it's it's a thing around the CPU. Then outside of that you have, as you already said, like the tools. And you could also and that I've recently been thinking about that a lot because I find it super interesting. If you s go one step further and think, okay, AI is solved and you you look at this the next big step and also what came out of the industrial revolution, I think it's really interesting to look at humanoid robots or embodied AI because the step when you leave the the virtual world that we are currently in and you say you want to do a lot more things, and at least to me it's really interesting. to look at the robotics space. Have you do you have any idea all around that or any feeling what will happen there?

Martin Tobias - Incisive Ventu:well I don't invest in robotics, or embodied a a AI, but I I think w what I'm actually working on a a deck and I'm doing a a presentation, ooh, I have to do that today. 24th, it'll be on my Twitter feed, but it is why you know B2B AI is is still a good investment in the era of the SaaSpocalypse with Salesforce going down or recovering and stuff like that. But anyway, but one of the slides that I have is the difference b w what is happening today versus what happened maybe 150 years ago. and what is happening today is a shift in where cognitive intellectual work happens, which is very similar to way may 150 or 200 years ago what happened between a shift in physical work from humans and animals to machines. If you go back to the time before we had Had lots of machines doing at agriculture and stuff like that. Muscle work, you know, work of moving things around was done by humans and animals, like 100% of it. but if you look today, about 100% of that mechanical work moving things from here to there is done by machines, cars, tractors, things like that, right? So if you looked at just the the the the muscle work that was done 200 years ago, it was all done by animals and humans, today it's ninety-nine per percent done by machines. And then you look at the the second thing that the the the bodies do, the cognitive work. that has you know very slightly been done by machines, adding machines, calculators, things like that ca have done a little bit of it. But what this AI thing has is is enabling to be done is a massive amount of that to be done by computers and that's the transition that we're in. Now some of those you know CPUs or digital brains or whatever you call it are going to be in chatbots and things like that. Some of them are going to be in harnesses because a robot is just another kind of harness right it is a harness that actually has a body of something either a humanoid robot or a physical robot. It's just in my mind robots are just a different kind harness that are designed for a different kind of for for for completing a different kind of thing and but they all need this digital brain that is better or are you know able to scale at a way that a human brain or an animal brain cannot And what I think the interesting opportunities are in robotics, I think there's going to be a lot of of very interesting growth in robotics, but what we will likely see and what we already have seen is the more of these specific function robots as the first ones that get to scale versus the general purpose. Everybody wants, you know, from the Jetsons the maid that can do anything. And maybe, you know, Elon produces that with his humanoid robot. But the problem you have there is training a general purpose humanoid robot to do everything you want to do is gonna take a really, really long time. I mean it might be a To do two or three things well out of the box, but it's gonna take a long time to be able to do everything in the house. And that's why what I think happens at robots is that you see the penetration first in things that are very specific, and you're already starting to see that. I mean, all these food delivery companies that have these little robots that drive around San Francisco, that that's doing one specific thing, it's a robot delivering food. I think you'll be able to get a robot delivering food really, really well and make that model work before you. you'll get a general purpose robot who can also deliver food. Do you see the difference? And it the same but but but more likely, I actually don't think that it's gonna be the consumer robotic versions, things it gets first. You you're already seeing you know companies like Amazon deploying robots in warehouses and getting those specific purpose warehouse robots to work at at scale and to r r replace humans at scale. It's most likely specific function industrial robots that get adopted and really change the economics of things quickly.

Niklas:And I think that's been happening. You could also argue that Waymo is a robot, right? At the end of the day, so it's an autonomous car

Martin Tobias - Incisive Ventu:Mm-hmm, mm-hmm.

Niklas:that drives around. So so I think we'll have it I think it it will be really interesting when we see hardware LLMs and robots being driven there. I also believe your side that there's a

Martin Tobias - Incisive Ventu:Yeah.

Niklas:lot of room and software actually. So if we look at it like a a harness would probably somehow work with proprietary data integrations and some kind of work. flow. If you look at what you're seeing, will it still be like web based, like a classical SARS? Is it running in the common line? What do you think where's the most interesting stuff happening currently?

Martin Tobias - Incisive Ventu:Well, I think the most interesting well I I've made two or three investments in in this area, so obviously I think they're interesting. I think the most interesting places for these new harnesses are in new new software for old industries and old industries that were somewhat that that were forced sort of reluctantly into the digital age and have fundamentally people who are not digital native but they have to have digital capabilities somehow. Let me give you two examples. the first one is Aura300 dot AI, which is one of my investments, and they are creating a voice agent and right now a a Yelp agent and a Google Ads agent for hair salons. It's basically the digital Back office AI agents to run a hair salon. And they're gonna be able to do more things. But basically, if you look at what a hair salon is, usually some lady loved doing hair for her friends, started a hair salon to do hair with her friends. She did not start it to answer the phones, to buy Google ads, to answer Yelp reviews, to do accounting. There's a whole bunch of stuff that you have to do to run that business reluctantly because your core business that you like doing is doing hair with your friends, right? So w they have been these reluctant technology adopters because their core business is not technology. yet they have to in order to do scheduling and text people and remind them and fill the seats and offer them shampoo and c and all these other things. So anyway, Aura 300 is building a whole suite of of agents that basically take all of that digital footprint workflow off of a human so the human can free up more time to do what they like to do, which is to do more hair. I think there's a hundred or thirty thousand different industries that that that are like that that are basic that have basically been reluctant technology buyers that will jump at and are jumping at the opportunity to have some digital assistant or agent or whatever you want to call it take all that digital shit off their plate because it's not core and strategic to the primary product that they have anyway, right? So I think that that that that's one. Another one that I have is a a company called Quintess AI, which is a voice agent for that also has an ERP for railroad and heavy equipment maintenance operators. And and this is another new software for old industry. So this is an old industry. These the average s maintenance tech in heavy equipment and railroads is like 60 years old, and their core job is Is wrenching with a wrench on a big rig or a railroad car and fixing something in the physical world that's broken. Now, where are they not digitally native but are required to? So at the end of the shift when they finish, they had to go back to an office and type in into an ERP system what they did with photos and say I fixed this and they update the ticket. And so there was a bunch of software that They had to use reluctantly, which was not their core business. And you know, it didn't really fit into the workflow to leave the shop and go do in an office an administrative task, which they hated doing. So now they have on their phone an app with a voice agent, they can say, Here's what I'm doing. They take a video of the thing, and so now they have integrated into their workflow in the field. something which is powered by AI which allows them to do their administrative job integrated better into their core job. And again what that is doing is making is software which makes this sort of administrative thing sort of disappear into the background and integrate better into the core thing that they want to do, which is wrenching on a machine to fix it. I think there is hundreds of industries for whom software is something that's required but there a little bit reluctant and those workflows and those harnesses are gonna get 10 times 100 times better with AI and disappear and enable those people to do more of what the core of their business is. And that is frankly as much as we sit here in Silicon Valley and think that LLMs and technology and software companies are the cool thing, the vast majority of the economy is not that.

Niklas:Yeah, and I think this is a huge opportunity to be honest. But what I thought about lately is and I I've built a small tool recently that bases on my thesis. So my PhD thesis was around production management. I dealt with early estimation of work content in the one-of-a-kind production. What you want to do is you want to estimate your product relatively early. to kind of then be better to estimate because you have a lot of uncertainty. And the clearer you are where your uncertainties are actually in your process, the better. So what I thought back then, it would be really nice if I could just talk to something and get an kind of estimate. And I think that is where we are moving. We are moving to this descriptive world. Before that, somebody would sit down on a C A D program, for example, and would construct whatever a ship. And then you would have a construction. Now we're moving one level up. So nobody's drawing the lines anymore. People are talking words to a machine, and the machine then does the drawing of the lines. They do the coordination of things in the background. That's what this harness does. And I think this is still, and this the really nice thing about what ChatGPT is actually doing is people are now used to describing things in a text box and just shooting that. So I think this is actually because of this market adoption of ChatGPT, for example, it enables these B2B markets where people will now be willing to actually type something in a text box and expect these things to be done that were done by an employee in former times. And I think this is a huge shift actually in the market and ESD also brings a huge opportunity.

Martin Tobias - Incisive Ventu:Yeah, I I agree. I mean the it's good to see that trans tr transition because when ChatGPT first started it was kind of like a a little bit of a toy because people were I remember showing my daughter who's in medical school now at Georgetown it for the first time, and it was like asking the smartest person you know or someone thought was pretty smart questions and getting and seeing if they knew things. It was you're trying to trick them to see if they knew something, right? So it was it was you were having you were asking a question to find an answer from the the smartest person you know. And what happened is with three, four, five and six, the person got smarter and smarter smarter. But luckily we didn't just go on the making the person smarter. We actually gave that gave that person more and more tools going from you know the either the CPU and the OS or the brain and the body. Now that chat that person that you're chatting with can do more than just give you an answer. It can research things it can open a browser and click buttons it can actually go and complete things, it can buy things for you, it can it it has a lot more tools. It's not just a Q and A bot. And the interesting thing that's happening now with sort of voice and robotics is that that that that thing that people now mate in a chat window ask to to to create whole work products instead of just answer questions. That whole work product could also now the the the the that experience could be gotten by having that same conversation with a robot or, you know, other form factors, and you know, through voice and not just through chat, which is the the really interesting thing that's happening in the in the evolution now is it's not just a chat box interface. It's a voice interface and soon a robot interface. And you already have some of those robot interfaces like you mentioned with Waymo and all of the the the the Tesla cyber cabs and stuff like that. You can just talk to them and they go where you want.

Niklas:On the other hand, while I think these are gigantic opportunities, we have the other hand where we just see some capex correct or some corrections in the general venture market. We see

Martin Tobias - Incisive Ventu:Yeah.

Niklas:kind of the value of Anthropic getting slashed roughly by fifty percent in a matter of a week. we are also seeing the first stronger discussions around this new cloud, and the question is this business model really sustainable? What's your take there?

Martin Tobias - Incisive Ventu:Well, the the the interesting thing, about a year and a half ago, I posted something which was you know, I I I said, you know, the we we are in a new transition and AI's gonna change everything, but it's not gonna be without bumps. And I d I definitely you know, the way that this build out is being financed is different than the way the internet and mobile build out was done so I don't think we're gonna have the explosion that we have, but I actually called out the weakest part of this AI story as the neo clouds. and the reason I did is that I think that that that leveraging everything with debt is i i i i i is a really risky way to do it. The the major the top four cloud providers are buying Buying all of this stuff off their balance sheet, and that's a better way to build infrastructure than to do it with debt. So I actually shorted Core Weave a year and a half ago. And the interesting thing is that it is starting to come around now that the that their debt is, I think, in junk bond territory, and it's because you know if you just look at a high level, what's the best way to build infrastructure? if you're building it 100% with debt, you've got to have contracts that are super solid and the that the and and you know forever and you just don't have that. You're starting to see know, Microsoft already lowered their commitment to Core Weave because they want to build more of their own data centers. And so I I I I I but I think that the the the the slowdown in in who's building these data centers, it's gonna switch from the neo cloud debt funded guys back. I think the total amount of data centers still gets built is a lot, but the question is who's building them? Is it Microsoft and and Google? Google, the people with big balance sheets or is it the neo clouds? I think it's gonna continue to be the people more with the big balance sheets. then to but but we're also gonna see a a a right sizing in the valuation of some of these things and that's what's happening with anthropic. I'm excited about the anthropic IPO. It's gonna unlock a lot of money from people that vent in you know LPs that invested in that venture to be able to reinvest in the next wave of things that is coming along. So I think it's gonna be an important liquidity thing for that for the overall venture market. and then the question of what's the right price, the fact that maybe it's half of you know today what it was A week ago. At the end of the day, it's going to happen at the at the valuation that it's going to happen. I think the prior discussions of it being a you know $2 trillion was probably too high. Frankly, I'm not a public markets investor, so I don't know what the right number is, but it will happen. And it's probably better for the venture ecosystem if it happens at you know $1 trillion or even just under $1 trillion, because then the public market investors will have a little bit more up. Side, one of the things that has actually been a negative for the for the venture ecosystem. Well, what used to be a positive was if you look back at the top the Mag 7, you look at all their IPOs. I mean, Amazon went public at a $700 million market cap. Microsoft went public, I don't know, like less than a billion. So, and if you look at the enterprise value of Google, 98% of the Enterprise value of Google has been created in the private in the public market. The same with Amazon, the same with Microsoft. And what's happened in the venture market in the last sort of 10 or 15 years is that 90, 95% of enterprise value has been created in the private market, and the public markets have basically been flat to down for these companies. And the public investors have not been getting the value creation because they've been buying on. Very, very late. And what you know helps the venture ecosystem in a weird way is if the public market investors can make money on these things because then they will be happier buying more of these IPOs. And frankly, I think anthropic going out at a more reasonable public market valuation is going to be good for the venture ecosystem because it'll mean demand. If they make money on that, there'll be more demand by the public market. investors for more of these kind of companies. If they go out at these crazy high valuations and end up crashing, this is what destroyed the SPAC market, you know, when Chamath screwed up the SPAC market, is that he took a lot of things at a spAC price that I think the SPAC IPO market is down like 85% or something like that. Like almost every company that went public on a SPAC is down. And when that happens, basically the public buyers are like, we don't want any more of that shit. And if all you get In the all you had coming to the public markets were these crazy high valuation private AI companies that ended up trading sideways or down after the IPO, the there would be very little demand for more of this. And I actually think a lower you know initial price for Anthropic is going to be good for the overall venture ecosystem because it might mean that public market buyers could make.

Niklas:So first of all, I would also say still say one trillion is a crazy evaluation in a sense, the achievable of the founders to get there, it's unbelievably outstanding. It's such a

Martin Tobias - Incisive Ventu:It's it's it's still incredibly positive. Yeah, yeah.

Niklas:great, great result, to be honest. Even even if now people are talking about slashing it, but if you look at where it was like a year ago, something like that, and where they are now, that the story is still crazy good. And I would also agree if they go public and l then the stock tanks in the next year and it loses fifty percent of market value, seventy five or whatever. That's not not a good sign for the overall market and probably not in the interest of enthropic also. Like they might win short term, they might get a little more capital, but long term I think it would rather be damaging. For them, it's probably better to have a success story in the public market after. So The easier the success story, they might actually benefit from it. So that that would be my my take as well. And once again, the result is amazing, like what from what they have have achieved also for their investors. If they go public at this valuation and you were early in that, like if you people were also big angels, were were really early and will mm, like Eric Schmidt, I think, for example, it's Really nice success story. And then again, we go back to the Neo Cloud. And I find this really interesting because Mag and Drazen talked about that. And he said, okay, it's what actually destroyed the dot-com bubble were the fiber companies that bought fiber at too high levels,

Martin Tobias - Incisive Ventu:Yeah.

Niklas:like on long-term contracts, and then the the prices or like the subscription fees actually went down so much that the business models weren't sustainable anymore. What he also said is. First of all, everybody after profited from it. obviously these companies got bust, somebody Yeah,

Martin Tobias - Incisive Ventu:Eventually it got used.

Niklas:but it got used. It's still the infrastructure that everything else, the whole internet got built on later on. So it it's still like the investors didn't profit, but other people profited from it largely and other companies as well. And for these new clouds, it might honestly be the same. It's hard to say. But if if GPU price is correct, for example. They are they locked in long term. This might happen. But what I also don't think is the compute demand will go up further, somebody else will make bank on it. There will still be the the overall market will continue growing. The revenue is real, this is not like Entropic and these companies. There's real usage on these things. These companies are making real money and other companies are also making real money. So I But think even if you see some part of the market folding, there are still really large parts of the market that will continue developing well. That would be my prediction.

Martin Tobias - Incisive Ventu:No, I I i absolutely because i it's part of my Presentation which I'm doing for the 24th of November, but if you look at most of the revenue from ChatGPT, well, ChatGPT is a lot got a lot more consumer, but if you look at anthropics revenue, most of it's from Claude Code. And so they have a great and amazing amount of revenue coming in. But if you figure like 70-80% of their revenue is from Claude Code, which is basically one harness, one market, one use case for this brain. and that one use case, there's 1.5 billion knowledge workers. There's only two percent of those knowledge workers who are programmers. So if you look at that, they have you know this amazing business, which is one of the fastest growing best businesses in the in the history of the world, pro you know, tra changing the way people do a put a a job that that's that's very important. It's a programmer, but at the end of the day, that's like two percent. of the market opportunity for this brain. and all of those other prof you know knowledge workers workflows are different harnesses that have to be built and that are being built. Harvey's building one in in legal, there are other people building them all along. But you know the vast majority of this amazing growth and amazing business that we've seen is on the back of literally like one percent market penetration. In knowledge work. Which to me says there's there's just tons and tons of opportunity and demand is gonna continue to outstrip supply. I I don't think that the demand and the cost for compute's gonna go down. in fact you probably saw SpaceX you know, just signed I think the most expensive compute contract ever with Google. and Google paid it because they could make the math work. They c they could afford to pay the compute is so short that they could afford to pay it. And you know, as we get more compute, I I still think demand is well ahead of compute supply for at least the next year and a half and we'll see in in a in a year, is my take.

Niklas:Currently it's definitely true. So for a week ago, there was a paper published on self-healing LLMs. And I haven't really looked at it, but it's a frontier LLM editing weights of a smaller LLM directly. And I was just curious. So I I went up, I spun up a TPU, took a look. I haven't done that in ages. It was more expensive to rent actually than a year ago when I first

Martin Tobias - Incisive Ventu:yeah.

Niklas:looked at it. So like same model. So currently it's definitely that's the case. So do you have currently the market is definitely rather we are short on compute and the demand is higher than it was than it was a year ago for the same hardware that you could also rent a year ago. So this is I think this is really really interesting and I would agree that it's not in the foreseeable future Not going to change. What I'm very interested in is really open weight against not necessarily the the the absolute frontier, like fable, but a level below. Because I've built a a piece of software that you can use to describe what you want built, like a table, for example. It will generate a 3D model. You can get a build of material and then you can build it yourself. And for that I use Chinese open weight models, and they are just like Ten they cost 10% roughly of what I would pay if I went to Entropic.

Martin Tobias - Incisive Ventu:Yeah.

Niklas:And this is really this is an interesting case because if you optimize, you will find something that is like that costs 10% of what Entropic is charging actually, with at a similar capability. Not at the absolute frontier, but they are good enough.

Martin Tobias - Incisive Ventu:absolutely. And you probably saw I think I posted this on Twitter maybe a month or two ago and that was the open router graphic that showed in the last three months the total number of tokens has switched from going, you know, majority to frontier models to now majority open weight models. The majority of tokens through open through through open router is now going to open models. And but the interesting thing is at the same time that that happened, you still have anthropic and open AI's revenue going through the roof. So I mean people still you know there are still enough use cases for the frontier to keep their revenue going. So I don't think that that I mean and and that just shows you again that demand for total tokens is is is through the roof and and The question is then, how much of those tokens are are frontier? you know, I think there's enough tokens to front even at twenty percent of total token usage, that's you know, a couple hundred billion of revenue for the frontier models still. and and every application company that I invest in, I tell them this the same thing, which is to look at all your calls like I did for investor match.pro, I have seventy two different calls. API calls for to to to to models. And I think you know when I first built it two years ago all those seventy-two calls were going to frontier models. But now it's something like only ten percent are going to frontier and ninety percent are going to second tier models just for cost and load balancing. And every application developer is gonna end up doing the same thing, saying, you know, what are some of the less complicated things that I could do with cheaper models? You you have to do that to make the cost of your a application work. But even by doing that, you're still giving plenty of revenue to the frontier models. yeah.

Niklas:Yeah, I think the same so and even even if we see this development continue, I mean there is enough space and I would say and value to be captured at the frontier. Like if you I think we see Entropic and OpenAI leaning a lot into science because it's actually hard and it justifies the prices, but it's also it's a really large market if you if you're just trained. Even if other parts of the market get captured by by other players, it's just large enough. I I would agree to that. there is just enough space and value

Martin Tobias - Incisive Ventu:Yep.

Niklas:to provide.

Martin Tobias - Incisive Ventu:Science, programming. I I think the majority of people doing programming want to stay, as close to the frontier as they can. there are other use cases, but but then there are other things answering your phone at your hair salon does not have to be the frontier.

Niklas:And the it's also the question if then Entropic and Open and AI are interested in moving into that space. And I think that's also always important to consider if you're working on something. if it's the obvious next thing Entropic and Open AI can do, then you have to be really quick and have a lot of money to somehow survive for a few days. Because it's not like if you yeah.

Martin Tobias - Incisive Ventu:Well that well that goes back to and I think I posted something about this on on Twitter as well. that goes back to the difference between a a a wrapper and a harness. you know and Anthropic published a report about three months ago that said that listed the ten most at risk jobs for AI. Number one being a programmer, number two being a customer service agent, number three being a salesperson. Anyway, they listed the ten jobs they think are most at risk or that have a percentage of a of of AI that could replace their work. And if you look at that list of 10 jobs, that's basically the 10 products that they're going to ship as a core harness from the model itself, the first one being Claude Code. I mean, they just told you what they're going to do. And so that's one reason why I never, when I I sort of knew this sort of before they published that report, but I did not invest in Replit, Bolt.new, you know, any of these coding things. Before Claude Co. because my thesis was that they that coding was so close to the model that they were going to do it well themselves. And in fact they did, and it's I and I think those opportunities are done. So they have already mentioned you know publicly, here's the 10 jobs. You can expect them to ship a harness on each of those ten jobs because they are very horizontal. Programming is something every company does, and it's a a very horizontal thing. So what you need to do is to find a harness in a vertical that is not one of those 10, but is still big enough to create a venture scale outcome. And and that is the the the the job of venture capitalists today is to find those markets and that that are big enough and and have enough integrations the the the the the other way to do it is to do what somebody like harvey did which is to Start a little bit as a rapper. They were rapping claud, but then get enough understanding of the legal workflows and everything where you create your own model. And that's exactly what Cursor did, right? Cursor started as a rapper and then created their own model and then ended up getting bought by Grok and became part of that. but but the the harnesses that are the most successful, many of them end up creating their own model because they they specifically train a model to be. Really good at that vertical. And a specifically trained model is going to be better than a general purpose model in a vertical that's big enough to afford its own model. I don't think every vertical ends up getting their own model, but it certainly gets enough of their own workflows and enough of their own integrations. I mean, in in you take hair salons, there's 28 different POS systems. It's not just like toast and stripe. And Anthropic is not going to write 28 POS integrations. They're gonna write toast, they're gonna write stripe, they're not gonna have all of the ones you need to do well in a sp in that specific vertical. So I think if you're building something, you need to be aware of what those 10 jobs are that the it models are very close to. Don't build in those and but build in one that's big enough and complex enough to have a a real

Niklas:Thanks. Last question, Martin. Which vertical would you be most excited to see pitches in currently? What what's it that that you find really interesting?

Martin Tobias - Incisive Ventu:that's a good question and it's hard to answer because I don't know it until I see it. but w w what I would like To see the the kind of pitch that I love is a founder who's been working in an industry trying to solve a problem, buying every software solution he could possibly do, and he has not been able to find anyone else to build the solution. And and he then has to build it himself. So I want somebody that's been working on a problem for five to ten years, has tried to Buy everything to solve it and hasn't found a solution, therefore, he has to build it himself. And hopefully, what they've done is run into like their college roommate who is a CTO, or maybe one guy's a business guy and the other guy's a technical person, but they have a some combination of okay, now is this is the problem I want to solve, and then I have a technical person that has can you can harness a new some of the features of this new. AI infrastructure to solve that old problem in a new way. So that's what I want to see is somebody that really understands the customer problem and is taking advantage of something in this AI technology that allows a new approach to an old problem that they know very well. That's the founder I want to meet. Founders.

Niklas:Thanks, Martin. It was awesome. See you in four weeks. And to you, my listener, also see you next time.

Download raw transcript (.txt) ↗

Listen to this episode

Open this episode on your platform of choice — or subscribe to the show while you're there.

More episodes