Latest / The Learning Corner by Precursor / Episode #10: Euclid Insights, Benedict Evans

Transcript

Episode #10: Euclid Insights, Benedict Evans · The Learning Corner by Precursor

0:00
The question is, is ChatGPT a good business?
0:02
I suspect it is.
0:03
In the sense that consuming $20 worth of compute for the average person, they're
0:08
probably okay with it.
0:09
But I also wonder, is this also like Uber?
0:12
Where Uber was great when I think it anywhere in San Francisco for $5 because my Uber ride
0:16
was subsidized by Uber and the VCs who backed the company.
0:19
Ultimately, it turned out to be pretty good because it created a transportation habit that
0:23
didn't exist before, but it took private markets financing a lot of that to get through
0:28
that curve.
0:30
It's the week of November 25th.
0:32
Welcome to the Learning Corner, a Precursor Ventures podcast.
0:38
Welcome to the Learning Corner by Precursor, and happy Thanksgiving week.
0:42
I'm Mia Farnham, and I'm joined by our GP and managing partner Charles Hudson, where we'll be
0:47
talking through the articles and good reads that caught our team's attention over the past
0:52
week.
0:52
Our first topic is from Euclid Insights titled The Siren Song of Services.
0:58
This piece delves into the concept of service as a software and how it differs from the
1:03
traditional software as a service SaaS model.
1:06
Unlike SaaS, which focuses on delivering scalable, self serve software products with
1:11
reoccurring revenue streams.
1:12
Service as a software involves offering services powered by technology, but far more
1:17
reliant on human operations.
1:19
These services often reach a larger TAM with highly personalized services, but frequently
1:24
lack the scalability and margins that true SaaS models have traditionally benefited from.
1:29
Euclid warns that the allure of service as a software stems from its ability to generate
1:34
early cash flow and foster relationships.
1:37
However, this approach can lead startups down a path where they are effectively building a
1:41
consulting business rather than a scalable product.
1:44
As the article sharply puts it, service as a software is the comforting lie that you can
1:50
build a software company while running a service business.
1:53
The piece highlights a key tension.
1:55
While services can help validate product market fit and generate revenue in the early days,
1:59
they often create operational complexity, detract from core product focus, and make it
2:03
harder to achieve the scalability VCs look for.
2:07
Seems like in this time of AI, services as a software companies are perhaps excited to win
2:12
back some margin through embedding capabilities across the platform.
2:15
Euclid states that companies should still tread lightly here since we haven't really quite
2:19
worked out the extent and depth of capabilities.
2:21
They gave the example of Klarna cleaning that they can reduce 2 thirds of customer support
2:25
costs through a partnership with OpenAI.
2:27
I really doubt it.
2:28
But wanted to know, Charles, what's your view on services as a software compared to
2:32
traditional software as a service models?
2:34
And I'd love to hear that perspective, especially as it relates to learnings from the
2:38
precursor portfolio.
2:39
Yeah.
2:39
I mean, the Youpoof folks making another appearance.
2:42
I enjoy reading their writing.
2:43
I think this all goes back to Sarah Benchmark's original post about service as a software.
2:48
And in some ways, I think the argument makes a lot of sense.
2:52
And in some ways, I think it's a reminder to how good a business model software as a service
2:57
has been for venture capitalists.
2:58
I think many of us are loathed to give up this high margin, recurring, easy to model, easy to
3:03
forecast revenue stream because it's maybe one of the greatest business models in the history
3:08
of time.
3:08
Sometimes it's hard to give up things that you love.
3:11
When I think about it though, I keep coming back to a couple of questions.
3:15
One is, as you noted, SaaS was a great business model because it was recurring, easy to
3:19
forecast, and high margin.
3:21
And if I think about this service as a software, one of my concerns is that we're
3:27
trying to, like, bridge from the known and safe to the unknown and safe using a familiar model
3:32
as opposed to, like, asking what are other ways that this could play out?
3:35
So I think I would be like, oh, we're gonna sell outcomes.
3:38
We're gonna sell the work.
3:39
And I don't know, Mia.
3:40
Like, one of the things I've learned from working with folks on the customer side is it's
3:45
nice to be able to plan.
3:47
It's nice to be able to forecast what you're gonna spend on a piece of software.
3:50
And sort of paying for outcomes, I think, is very different than paying for seats.
3:53
Like, seats have all kinds of other issues and challenges, but one thing that's nice about it
3:57
is your CFO can say, alright.
3:58
We bought 10 seats of Salesforce.
4:00
I know exactly what those are gonna cost for the year.
4:02
And at the end of the year, we we can ask ourselves, was whatever we spent on those seats
4:06
worth it?
4:06
I think with selling the outcome and selling the work, you don't always know what you're
4:09
gonna end up spending.
4:11
So that's one.
4:11
And the second thing, and this is where I keep getting hung up.
4:14
I've met a bunch of people who are like, we're an AI SDR company.
4:18
I'm like, that's great.
4:18
People are like, well, these companies pay these SDRs 60, $80,000 a year to do
4:24
prospecting.
4:25
Our solution is going to eat into that spend.
4:27
I'm like, but you're probably not gonna replace $80,000 worth of human activity with $80,000
4:33
worth of AI generated SDR activity.
4:36
That's just unlikely to happen.
4:38
So the real question is like, well, when you replace the human, what do you, like, what do
4:42
you replace it with?
4:43
And do you replace $80,000 of human with 50,000 of AI spend, 20,000 of AI spend, 10,000 of AI
4:54
spend.
4:54
Like, what's the number?
4:56
And the thing I keep worrying about is if you have 50 people trying to replace an $80,000 SDR
5:03
using similar, but not the same, underlying tech stack prompts, etcetera, that sounds like
5:09
a recipe for dramatic price deflation.
5:12
And it sounds like a recipe for a race to the bottom if you're not careful.
5:17
And eventually, maybe you end up replacing an $80,000 human with $15,000 worth of software
5:25
that cost the service provider $10,000 to deliver.
5:29
So now you've got certainly have a service as a software business that sells outcomes, but the
5:34
cost of those outcomes and the revenue you make, it's like a worse business.
5:37
The top line is lower and the margins are worse, but maybe you make up for it by not
5:41
needing as many people to build it.
5:42
And if you can scale it across enough customers, etcetera, etcetera, etcetera.
5:46
This is the thing I keep coming back to is the reason most people would rather sell software
5:50
over services is software has a marginal cost of 0, and it has all these great attributes
5:54
that I won't get into here.
5:55
And services really don't have that.
5:57
If you think about services and software, it's kind of a natural way of saying services
6:01
enabled software.
6:02
And we've seen those businesses in the past pre AI.
6:04
And pre AI, they're okay businesses, but they're not nearly as good as pure software,
6:08
which is why so many VCs push companies to be a 100% software and push down services even
6:13
though you can't really do that.
6:15
So I remain really curious to see as we replace humans or human labor or other activities with
6:21
agents or AI systems, what does the margin profile look like for these businesses, And
6:27
what effect does competition have in eating away at those margins?
6:30
Any conversations that are happening within our own portfolio where maybe the original business
6:36
model was truly SaaS, and then now with the emergence of AI, now maybe founders are like,
6:41
well, now we can kind of scale up this really cool enablement that can be highly personalized
6:45
for a customer, and we can do it and win back some margins through x, y, and z.
6:49
Yes.
6:50
I think that the two interesting conversations I've been having with folks are like, hey.
6:54
We figured out a way to use AI for our internal processes to make ourselves more efficient.
6:59
Should we pass the savings on to the customer in the form of a better price?
7:03
Or should we capture these savings in the name of margin?
7:06
And depending on the company and depending on the competitive dynamic and depending on their
7:09
customer type, I've said the answer varies.
7:12
It really depends.
7:13
The second thing I've seen is a lot of AI native or AI first vendors have come into
7:18
market to compete with companies that we have that provide traditional SaaS.
7:21
And then in some cases, those companies have made bold claims.
7:24
We're gonna give you 10 x improvement over what you have.
7:27
And they're and our companies have a wow.
7:28
If it's really 10 x, we can't we can't match that with what we have.
7:31
And I'm like, alright.
7:32
Well, let's see what the customer actually gets.
7:34
And in many cases, the customer is not getting anything remotely close to what's promised, but
7:38
they are getting something significant.
7:40
And the question is, if you're at 2 x, what happens faster?
7:43
Does the AI native company that promised 10 x go from 2 x to something more material sooner
7:50
than the company that's already got the customer is able to integrate AI and sort of
7:54
come down.
7:55
I think the argument in favor of the new companies is it's hard to cannibalize the
7:58
business you already have if you're selling software.
8:01
The the other thing is, though, if you're the the startup, you're kind of gated by how
8:05
quickly the AI models get good enough to allow you to improve your process.
8:08
And I'm actually not sure where those two curves intersect.
8:11
From your perspective, what advice would you maybe give to early stage founders who are
8:14
trying to strike the right balance between leveraging services and staying focused on
8:18
something that's more scalable from a product development stage?
8:21
I felt like I've had a lot of conversations with awesome founders who have really good
8:25
product, and then you kind of look at the back end and you realize, oh, it's because they're
8:28
running a dev shop.
8:29
So no wonder their product is so good, but how can you possibly scale that at the burn that
8:34
you're running it?
8:35
I couldn't agree with you more.
8:37
I think services are a great way to get to know your customer.
8:40
I love how we've, like, essentially rebranded pilots as design partners.
8:45
You know, the nomenclature changes, but I think this notion that sometimes solving the
8:48
customer's problem involves a combination of a software product and a service.
8:53
And for me, the question is always like, okay.
8:54
Does offering them services give you insight into things you can solve in product in the
8:59
future, even if you can't solve them today?
9:01
And if you think that providing services is a meaningful way to get insight into the customer
9:06
problem and make your product better, by all means, do it, and ideally, you'll get paid for
9:10
it.
9:10
But in some cases, you're just like, wow.
9:12
This services component isn't going away.
9:14
It's not something we can automate or make work.
9:16
It's just something we have to do.
9:18
I just think about for example, there were a lot of companies before that would give you
9:21
kind of white glove onboarding, which is a fancy way to rebrand services and, like, hey.
9:25
You give us all your messy data, and we will make it clean and get into the product for you.
9:29
That was a service that was unlikely to get materially better over time because everybody's
9:33
data was messy.
9:34
We might be in a world now where you can do white glove onboarding using using AI and
9:39
actually get better in your product at onboarding people's data.
9:40
So I don't think we should be as afraid of services as we were before.
9:41
But, also, I think it requires being honest about, hey.
9:43
Are you really building a product that where the services are informing what to build versus
9:47
just a part of the product delivery that you're gonna have to have forever?
9:54
Our next topic is the presentation created by Benedict Evans titled AI Eats the World, which
10:00
is his annual presentation exploring macro and strategic trends across the tech industry.
10:05
If you haven't opened up his presentation yourself, definitely take some time to do so.
10:09
We'll do our best to summarize the takeaways on the pod, but there's something really
10:13
entertaining about seeing it in visual form, in slide format.
10:16
So Benedicte walked through the surge in interest and use case for AI across nearly
10:21
every industry while calling out the fact that we are still working out the questions about
10:25
the space itself.
10:27
1, how far would this scale?
10:29
2, how is this useful?
10:31
And 3, how do we deploy this?
10:33
And while capex is surging across the largest foundational LLMs and capital floods to the
10:39
market to seize the opportunity, Benedict says, quote, everything is still changing under our
10:45
feet.
10:45
So Benedict calls out the fact that the consumer Internet model was to launch free, go
10:51
viral, and work out revenue later.
10:53
That clearly does not work with the LLM cost model.
10:57
OpenAI expects a loss of 5,000,000,000 on 3,700,000,000 in revenue.
11:01
Computing costs alone will reach around $6,000,000,000 this year, according to a report
11:06
by Foundational Capital.
11:08
So they definitely didn't launch for free.
11:09
And on the morality piece, we in our VC bubble can believe things like chat GPT have gone
11:15
viral while there's huge interest in generative AI.
11:18
There's actually limited use so far if you look at the percentage of people who use it on a
11:23
daily access point across the globe.
11:26
So how long are we going to live in this gap of spending time, money, and resources building
11:31
and still not creating something that is exactly what the consumer needs?
11:35
Boy, I have so much to say.
11:37
I always enjoy Benedict's presentations.
11:39
I I think the visual is really helpful, and there's some slides in there where I, like,
11:42
literally chuckled out loud when I clicked on them.
11:44
They were just with that right level of, like, pithy humor.
11:47
I think part of the challenge is everybody's struggling for an analogy for AI.
11:51
Is it like Internet 1 point o?
11:53
Is it like SaaS?
11:54
Is it like electricity?
11:56
And I'm like, I don't know.
11:58
I don't know if there's an analogy because it is certainly it feels to me like revolutionary
12:02
technology on the scale of, like, electricity.
12:05
But the rate at which the technology is compounding feels I wasn't around when
12:08
electricity was.
12:09
I mean, I'm old.
12:09
I'm not that old.
12:10
I wasn't around when electricity was invented, so I don't actually know what it felt like when
12:13
electricity was invented.
12:14
My sense is AI is compounding faster than electricity.
12:18
And, also, AI is gonna touch even compared to the Internet, I have a feeling that's gonna
12:22
touch more elements of life faster.
12:25
But I think if you settle on an analogy, it kinda dictates what you think is gonna happen
12:30
next.
12:30
So if you're like, oh, AI is like fiber.
12:33
It's like the fiber build out for the Internet.
12:35
We laid all this fiber, and it felt like it was ruinous, and who's gonna use it?
12:39
And lo and behold, we used it all.
12:41
And maybe it wasn't all efficient capex in the moment, but it enabled stuff in the future that
12:46
was highly valuable.
12:48
Or is it like the consumer Internet where, like, it is about being one of the early
12:52
winners and building a brand.
12:54
And I don't know what the right analogy is, but what I see a lot of people doing is trying to
12:59
fit their expectations for where AI will go to some previous I want, like, what if it's just
13:03
its own thing?
13:04
What if it's just really different and there isn't a really good historical narrative or
13:09
proxy?
13:09
There are a couple of things I think about though.
13:11
ChatGPT at $20 a month has kind of set the I don't think it's an accident that ChatGPT
13:17
clawed perplexity in just about any consumer AI product you use is $20 a month.
13:22
I think it would be pretty hard to charge $50 a month for a consumer product now because the
13:26
reference price is 20.
13:28
For me, the question is, like, is ChatGPT a good business?
13:31
I suspect it is.
13:32
In the sense that consuming $20 worth of compute for the average person, unless they're
13:38
really using ChatGPT a lot, they're probably okay with it.
13:41
There's probably an argument that there should be a ChatGPT Lite for 5 or $10.
13:45
But I also wonder, like, is this also like Uber, where Uber was great when I could get
13:50
anywhere in San Francisco for $5 because my Uber ride was subsidized by Uber and the VCs
13:55
who backed the company.
13:55
And, look, that ultimately turned out to be pretty good because it created a transportation
14:00
habit that didn't exist before, but it took private markets financing a lot of that to get
14:05
through that curve?
14:06
Or is it like pets.com, and you're shipping bags of dog food through a pretty brittle
14:12
supply chain at a price that's contribution margin negative the second it leaves the
14:16
warehouse?
14:16
So for me, the real question is, so long as the big LLM providers have the support and faith of
14:22
private capital markets, I don't wanna say it doesn't matter, but they have time to figure
14:26
all of this stuff out because they will be able to continue to foot the bill.
14:30
And my belief in having talked to these folks is they all understand this flywheel, which is
14:34
if they can make compute cheaper, people will use AI to do more stuff.
14:38
And they can, if they get it right, drop the cost of compute faster than they drop the price
14:44
to the end customer and still win.
14:45
But that basically requires a deep faith in your ability to access, my guess is
14:51
conservatively, another 10 to $30,000,000,000 in private capital for the next couple of years
14:56
to continue to make these investments, which is kind of a scary value prop.
15:00
That's a lot of money to raise.
15:01
I happen to believe that OpenAI, Anthropic, and a few other companies will have access to it.
15:05
The other thing is, I think customers are still trying to figure out what's all this stuff
15:09
worth.
15:10
I think right now, a lot of people are like, sure, we'll spend $1,000,000 on a pilot to try
15:14
a bunch of things with OpenAI or with Claude or with whatever model of llama 3, whatever you
15:19
wanna use.
15:19
But I think ultimately, at some point, maybe it's next year, maybe it's year after, the CFOs
15:24
of these companies is gonna come and say, we've spent all of this money.
15:27
What are we getting for this?
15:29
Are we getting value from this?
15:31
And right now, I think it's the other end of the spectrum, which is like we cannot be
15:34
perceived by our investors or by the market as not taking AI seriously and not making
15:38
investments.
15:39
So we need to be deploying resources.
15:41
But that pendulum at some point will switch, and people will say, but we also need to be
15:45
getting value from this.
15:46
And I just don't think we're there yet.
15:47
And when I read Benedict's deck, I was like, there are just so many open questions that he
15:51
posed that I was like, I feel so much better about myself because I also have these
15:55
questions and don't have an answer.
15:57
But I think many people I know, their view on where I is headed is based on, I'd say, a more
16:02
confident view of where things are headed than I personally have.
16:06
I think it was his last year's presentation, which also was centered around AI.
16:09
He had one sign that said, this is clever.
16:11
So what?
16:12
A personal example that's just funny, I, as the tech VC, in my wedding planning era, I decided
16:19
to cut down on questions being asked by my parents about what's going on with the wedding
16:23
weekend.
16:24
I created my own model within chattypt4 0.
16:27
Mhmm.
16:28
And I pummeled it with months worth of communication, emails, documentation on every
16:32
single thing that's going on in my wedding to pump out a model that I could pass over to my
16:36
parents.
16:37
And so if they have any questions, ask model, don't ask me.
16:40
And I probably spent, I wanna say 15 plus hours pummeling it with every single thing, feeding
16:46
the model.
16:46
You should've seen me when I had this done.
16:48
Was like, this is literally the world's most amazing complete model.
16:51
It's gonna knock everyone's socks off.
16:52
It's gonna be curing my problem.
16:54
I mean, this goes to the gapping question of, like, so what?
16:57
I pass it along to my parents, and I had one realization.
16:59
One, I couldn't even share my 4 o model because it had too many attachments and images there,
17:04
so they couldn't even use the model.
17:06
And then the second of which my sweet parents looked at it and they're like, oh my gosh.
17:10
Looks so great.
17:11
I'm not using this.
17:12
I'm going direct to you.
17:13
Like, what's the point?
17:14
And so I ended up just making a lot of screen record sessions of me asking the model, fix.
17:19
We're sending it to them.
17:21
Which they just thumbs up in my text message as opposed to actually absorbing it.
17:26
So it goes back to my question of even on a consumer basis, the people who are using it for
17:30
$20 a month, we still haven't quite hit it of this is clever, so what?
17:35
What are we using for?
17:36
With folks who are debating, spending 1,000,000 of dollars a year on pilots, on contracts to
17:41
use it to actually conduct business that they generate revenue on, we're we're far from it, I
17:46
feel.
17:46
Yeah.
17:46
I read a great article in the information this week, which was that, you know, a lot of the
17:51
model providers are like, oh, maybe reasoning is the big unlock for business value.
17:55
Maybe the pre reasoning versions of these models were kinda brute force.
17:58
And I'm like, if I were in the model business, I would certainly make that argument too.
18:02
And having used o one, I think it's great.
18:04
And so part of me feels like there probably won't be some big bang moment where we're like,
18:08
AI is good enough.
18:09
It's gonna be incremental incremental incremental.
18:11
And some people are gonna be like, it's it's good enough.
18:13
It's good enough.
18:14
And I think right now, I talk to people who are building things with agents, and a lot of them
18:18
are like, agents are getting better.
18:19
I'm like, getting better and good enough for production are actually 2 different standards.
18:24
And so this is one of those cases where I think the people who are on the front lines building
18:28
probably have the best perspective.
18:30
But like UBI, I've built some custom GPTs that I've deployed, some of them with more success
18:36
than others.
18:36
I will also just say it's important to continue to use these tools.
18:40
I think about one tool where, like, my opinion has changed quite a bit in the last 2 months is
18:45
actually perplexity.
18:46
When I used perplexity a few months ago, I was, like, just doesn't feel that different than
18:50
Google in terms of the results I'm getting and the reasoning.
18:52
And lately, as I've played with it more, I'm like, oh, I am starting to get a feel for
18:57
what's different.
18:58
But again, at $20 a month, paying for consumer search engine is a new behavior.
19:01
And so I'm very curious to see what existing behaviors change in a world of AI and which one
19:08
stay the same.
19:08
But $20 for a consumer is a lot of money unless you're using ChatGPT to help you with your
19:13
homework or some other use case.
19:14
I don't know that most consumers have $20 worth of queries and questions where ChatGPT or any
19:20
LLM is well suited to provide context and answers above and beyond what they can get from
19:26
traditional search.
19:27
Well, we are all out of time today.
19:28
Thank you to our listeners.
19:30
Please like and share with friends, and we will all see you next week.
19:35
Thank you.
19:35
Bye.
19:36
Bye.