November 17, 2025 • 23 mins
World Labs Unveils Marble: Transform Text and Media into Editable 3D Environments
Cursor Raises $2.3 Billion, Valuation Soars to $29.3 Billion
The 9-Month Rule or Why AI Startups Can't Outrun the Wave
#AI, #startups, #3Dtechnology, #innovation, #valuation, #entrepreneurship, #technews
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to Innovation Pulse, your quick, no-nonsense update covering the latest in startups and
(00:09):
entrepreneurship news. First, we will cover the latest news. World Labs unveils Marble,
a platform for creating editable 3D environments, and Cursor raises $2.3 billion to enhance its AI
coding tool. After this, we'll dive deep into the challenges AI startups face, exploring why
(00:32):
many struggle and what makes others succeed in a rapidly evolving tech landscape. World Labs,
founded by AI leader Fei-Fei Li, has launched Marble, an innovative product that transforms text
prompts, photos, videos, 3D layouts, or panoramas into editable and downloadable 3D environments.
(00:55):
Unlike its competitors, Marble offers persistent 3D worlds rather than on-the-fly generation,
enabling users to export worlds as Gaussian, Splats, Meshes, or Videos. It stands out with AI
native editing tools and a hybrid 3D editor that lets users outline spatial structures before
(01:17):
detailing them visually. Marble's flexible input options allow users to upload multiple images
or clips, creating realistic digital twins. Its chisel feature separates structure from style,
offering direct object manipulation, enhancing creative control. The product also allows world
(01:39):
expansion and composer mode for large spaces, making it ideal for gaming, VFX, and VR projects.
Available in four subscription tiers, Marble provides a range of features for various
creative needs. It holds potential for robotics training environments and aims to pioneer spatial
(02:00):
intelligence, a crucial step towards intelligent machines as envisioned by Fei-Fei Li.
Cursor, an AI startup, recently secured a significant $2,300,000,000 funding round,
boosting its valuation to $29,300,000,000. The company offers an AI coding tool designed to
(02:25):
assist developers in generating, editing, and reviewing code efficiently. Originally launched
within Anisphere, a research lab established in 2022, Cursor has quickly become a leading name among
AI startups, valued over $10 billion. The recent funding saw participation from major investors
(02:46):
like Axel, Thrive Capital, Andreessen Horowitz, and others. With this investment, Cursor aims to
deepen its research and enhance its product offerings. The tool has achieved a remarkable
milestone, surpassing $1 billion in annualized revenue and expanding to over 300 employees.
(03:08):
Notably, Nvidia CEO Jensen Huang praised Cursor as his top enterprise AI service.
Despite industry interest, Cursor remains focused on growth rather than an immediate IPO.
The coding tool market is competitive, but Cursor distinguishes itself with proprietary models,
(03:28):
generating more code than most large language models. Competitors include OpenAI and Anthropic,
each developing their AI coding tools, though Cursor continues to lead with its unique capabilities.
And now, pivot our discussion towards the main entrepreneurship topic.
(03:53):
Hey everyone, Donna here on Innovation Pulse, and today I've got someone who's going to make every
AI startup founder either panic or pivot. Yakov Lasker is a senior software development manager
who's been watching the AI startup landscape with what I can only describe as well-informed horror.
Yakov, welcome to the show. Thanks Donna. And yeah, horror is the right word.
(04:16):
You know what keeps me up at night? I just read that 90% of AI startups fail within their first
year. 90%. But here's the kicker, everyone's blaming it on the wrong thing. Okay, you've got
my attention. What's the wrong thing? People keep saying it's because the big tech companies are
more agile than we thought. Right. That OpenAI and Google and Anthropic are just moving faster than
(04:41):
traditional incumbents. But that's missing the entire point. So what is the point? The ground is
literally shifting under everyone's feet every nine to 12 months. It's not that the big players
are outrunning the startups. It's that they're the ones causing the earthquakes that keep knocking
everyone else over. Wait, hold up. Nine to 12 months? That's insane. I mean, most startups need at
(05:06):
least 18 months just to find product market fit, right? Exactly. And that's the trap. Look, I came
across this investment thesis that basically argues every AI application startup is doomed.
And the data's pretty damning. In 2024 alone, ChatGPT went from 200 million weekly users to
(05:28):
400 million by February the 2025. The mobile app generated over a billion dollars in just the first
part of this year. That's 53 times more revenue than its nearest competitor. Okay, but ChatGPT
is ChatGPT. Surely there's room for specialized apps that do specific things better. That's what
everyone thought until OpenAI's Dev Day in October to 2025. They announced they're turning
(05:54):
ChatGPT into an entire platform. Basically an operating system where other developers can build
apps inside ChatGPT itself. Oh, no. Oh, yes. They introduced something called the apps SDK and
Agent Kit. So now instead of users leaving ChatGPT to use your specialized AI tool,
(06:16):
they can just use a version of your tool that lives inside ChatGPT where they already spend all
their time. That's like if Amazon decided to just clone every successful product on their marketplace
and sell it themselves. Oh wait, they kind of do that. Exactly. But it gets worse. Remember all
those AI scheduling assistants. ChatGPT now does automated task scheduling. AI image editors,
(06:43):
ChatGPT has image generation and editing built in. Customer service bots, code completion tools,
search engines. They're just eating everything. Everything. And it's not just OpenAI.
Anthropic went from Claude Sonnais 3.5 in June 2024 to Claude Sonnais 3.7 in February 2025
(07:07):
to Claude Sonnais 4 and Opus 4 by May 2025. That's three major releases in 11 months.
Okay. So I'm hearing a lot of doom and gloom here. But surely some AI startups are making it
work? I mean, we keep hearing about billion dollar valuations. Sure, but here's the thing. Those
valuations are based on potential, not proven business models. There's this fascinating
(07:33):
stat from CB Insights. The 32 new AI unicorns in 2024 represented almost half of all new unicorns,
but they haven't built commercial networks nearly as robust as non-AI unicorns. They're
reaching unicorn status with 203 employees instead of 414 and in two years instead of nine.
(07:58):
So they're growing fast, but not deep. Right. And the money tells you everything. Venture funding
for AI startups dropped 42% in 2024. Most companies that got funded in the 2021 to 2023 boom had 18
to 36 months of runway, which means by late 2025 or early 2026, which is basically now,
(08:23):
they're running out of money. And they can't raise more because investors are getting smarter.
They're getting more skeptical for sure. But here's where it gets really interesting. The
big tech companies aren't even doing traditional acquisitions anymore. They're doing these reverse
acquihires. What's a reverse acquihire? So Google paid $2.7 billion for character AI,
(08:46):
but they didn't buy the company. They just hired the founders and key talent and licensed the
technology. Amazon did the same thing with a robotics AI startup called Cavarian. They took
the founders and 25% of the workforce. Wait, that's brilliant. You get all the talent and the tech
headings of a full acquisition. Exactly. Microsoft did it with inflection. And this is happening at
(09:11):
scale. In 2024, there were 384 AI acquisitions. And a ton of them followed this pattern.
So the two ways to win as an AI startup are what? Build fast, generate cash and get out?
Or build something good enough that Google acquires your team?
That's basically the thesis, yeah. You've got maybe 12 to 18 months to generate cash before
(09:35):
the foundation model providers add your feature to their platform. Or you build something impressive
enough that they want to absorb you. Okay, but I'm hearing a lot of people talk about moats.
Surely there are ways to build defensible AI businesses? There are, but they're very specific.
And this is where the thesis gets more nuanced. The survivors aren't the ones with the best AI
(09:57):
models. They're the ones with the stickiest integrations, the most unique data, or the
deepest vertical specialization. Give me an example. Take a bridge AI. They do medical transcription,
turning doctor patient conversations into clinical notes. But they're not just doing speech to text.
They understand clinical workflows. They've tuned their models on medical terminology.
(10:22):
They integrate with Epic, which is the dominant electronic health record system.
And they embed seamlessly into how doctors actually work. They're now in over 100 health
systems. So it's not about having better AI. It's about being impossible to rip out. Right.
Or take scale AI. They got a five year $100 million contract with the Department of Defense.
(10:48):
To compete with that, you'd need security clearances, SIF infrastructure,
and years of relationship building in government. That's a cornered resource moat. You literally
cannot copy it. Okay, so there's a pattern here. Medical, defense, what else?
Highly regulated industries. Places where you need domain expertise that takes years to build,
(11:09):
or hardware integration, like if you're putting actual devices in people's homes that learn and
optimize over time. The key insight is that the moat can't be the AI model itself, because everyone
has access to the same foundation models. So if you're building on top of Open AI's API,
or Anthropics Clawed, you don't really have a moat. Not a technical one. No. There's this brutal
(11:34):
medium article I saw titled zero moat in AI application startups. The author looked at over
2,000 startups and basically said, if you're just wrapping a nice interface around GPT-4,
you're building on someone else's infrastructure with no defensibility.
And when that someone decides to build the interface themselves, you're done. The article gave this
(11:57):
perfect example. Dozens of AI resume builders launched in 2023, then LinkedIn and Canva just
added the feature natively. All those startups, dead.
Okay, so this is starting to sound less like AI startups are doomed, and more like,
most AI startups are building the wrong thing.
(12:18):
That's a better framing, actually. But the original thesis has this insight that I think is
really important. It's not just about building the wrong thing. It's about the rate of change
being so fast that even if you build the right thing, the foundation shifts before you can mature
as a business. Explain that. Okay, so think about previous technology waves, the PC era,
(12:41):
the internet boom, mobile. Each of those took years to play out. You had time to build a company,
establish sales relationships, develop brand recognition, become an incumbent yourself.
Right. Like how Facebook started in dorm rooms and had time to become the giant that now buys or
crushes new social networks. Exactly. But with AI, the sea changes are happening every nine to
(13:06):
12 months, and very few startups can turn into a mature business in that timeframe. Even if your
engineers can pivot quickly, you can't hire 100 people in a month without your organization
imploding. Sales cycles, customer relationships, brand building, these things are incompressible.
So you never quite make it to the finish line before the race course changes.
(13:29):
That's the metaphor the thesis uses. You can't skate to where the puck is going if 20 people
are about to slap it in unpredictable directions at high velocity. I love that. Okay, but let's get
practical here. If you're an investor or a founder, what do you actually do with this information?
Well, if you're an investor, you're asking very different questions now. Not just how fast can
(13:53):
you grow? But how defensible is this specifically? The thesis suggests looking at whether the
fundamental assumptions underpinning a startup's existence will be the same in five years,
or whether they'll be unpredictably different. And if they're unpredictably different, that's
not investable. Not as a long term independent company. No, you might invest expecting a quick
(14:17):
flip or an acquisition, but you're not betting on the next Google. What about if you're a founder?
If you're a founder, you need to be brutally honest about your moat. Are you building something
that requires years of domain expertise? Do you have proprietary data that gets better with every
user interaction? Are you integrating so deeply into customer workflows that switching to a
(14:39):
competitor would be a nightmare? Do you have regulatory advantages or hardware in the real
world? And if the answer is no to all of those, then you're probably building a feature, not a
company, which might be fine if you can generate cash quickly, or if you're building to be acquired.
But don't fool yourself into thinking you're building the next sales force.
(15:00):
There's another angle here that we haven't talked about. What about open source?
Meta's releasing Lama models for free. Doesn't that change the equation?
It's interesting, right? The data shows that while open source is advancing, there's actually a
consolidation happening around the closed source frontier models. Anthropic now has 32% of the
(15:23):
enterprise market. Open AI has 25% down from 50% two years ago, but Meta's Lama only 9%.
Why is that? I would think free and customizable would win an enterprise.
Two reasons. First, the closed source models are still meaningfully better for most use cases.
(15:44):
Second, and this is crucial, enterprises care about reliability and support.
If something breaks in production and you're using Lama, you're on your own. If you're paying
Anthropic or open AI, you've got a throat to choke, as they say in enterprise sales.
Plus legal liability, compliance, all that fun stuff.
(16:05):
Exactly. Though there is this interesting dynamic where some developers use open source for
experimentation and then switch to closed models for production. But that hasn't translated into
a real business model for the open source model providers. Like, Mistral AI has an unclear path
to revenue according to the information. Less than 10% of users pay for their commercial models.
(16:32):
So open source is great for the ecosystem, but not necessarily for building a business.
Not at the foundation model level anyway. Though some startups are building businesses
by fine tuning open source models for specific verticals. That's a different play.
Okay, I want to circle back to something you mentioned earlier. You said there's a pattern
(16:52):
of successful AI startups in healthcare, defense, regulated industries. Is there anything else that's
working? Yeah, this is where we get into what the thesis calls real atoms data. Basically,
if your AI is dealing with the physical world, hardware, manufacturing, logistics, agriculture,
you might have a shot because that data is much harder to replicate.
(17:16):
Because you can't just scrape it off the internet.
Right. And it often requires sensors, equipment, physical presence. There's a startup called
Goodship that does freight management. They're not just predicting stuff, they're integrated into
the actual operational workflow of moving physical goods around. Replacing them would mean retooling
(17:37):
mission critical logistics infrastructure. So the pattern is go deep, go vertical, go physical,
or go regulated. That's a good summary. And ideally, combine multiple modes. Like that
medical example with average, they've got regulatory modes, workflow integration,
proprietary data that gets better over time, and domain expertise. Any one of those might not be
(18:02):
enough. But all four together, they'd be really hard to kill. Exactly. Though even they're not
completely safe. If open AI decided to partner directly with Epic and built medical transcription
into the EHR. Oh, that's terrifying. Welcome to being an AI application startup in 2025.
(18:22):
Okay, so let's talk about what happens next. We've established that most AI startups are in
trouble, that the foundation keeps shifting, that there are some defensible positions.
What's your prediction for the next 12 to 18 months?
I think we're going to see a massive wave of consolidation.
Bessemer Venture Partners is predicting a surge in M&A activity in 2025 and 2026.
(18:47):
The AI native startups that pushed into specific verticals will get bought by traditional
software companies that need to evolve or die. So like Salesforce buying a bunch of AI startups?
Exactly. Salesforce is already buying Informatica for $8 billion. You'll see more of that.
The traditional SaaS companies realize their products are at risk of being AI-ed out of
(19:10):
existence. So they're buying capabilities rather than building them. Because building from scratch
would take too long. And by the time they finished, the technology would have moved on.
So you're going to see the giants buying the upstarts that actually figured out how to apply AI
in specific domains. The startups get liquidity, the acquirers get to survive.
(19:32):
What happens to everyone else? The ones without moats? They run out of money and shut down.
We're already seeing this. That MIT study showing 95% of business AI implementations failing?
Those aren't just pilots. Those are startups dying because their product got commoditized
before they could scale. And the foundation model providers just keep expanding.
(19:55):
Yeah. They're in this weird position where they're simultaneously infrastructure providers
and application builders. OpenAI is a great example. They provide the API that powers a
thousand startups, but they're also building chat GPT into a platform that competes with all those
startups. That seems like a conflict of interest. It's complicated. From their perspective, they're
(20:19):
just adding features that users want. From the startup perspective, they're pulling the rug out.
And because the foundation model providers control the underlying technology,
they can move faster than anyone building on top of them. So if you're building on OpenAI's API,
OpenAI can always out-execute you because they see what's working and they control the foundation.
(20:42):
In theory, yes. Though I think the reality is messier than that. OpenAI is building for
horizontal use cases. If you're deep in a vertical, like medical imaging or legal discovery,
you might still have an advantage because you understand the problem space better.
But you're always one product announcement away from being obsolete.
(21:03):
That's the existential dread. Yeah.
All right. Let's bring this home. If someone's listening to this and they're thinking about
starting an AI company or investing in one, what's the one thing they need to internalize?
The foundation is not stable. This is not like previous technology waves where you could ride
a trend for five or 10 years. The ground is shifting every nine to 12 months, and it's
(21:26):
shifting in ways you cannot predict with confidence. So whatever you're building,
it needs to be defensible not against competition, but against obsolescence.
And if it's not defensible, it needs to generate cash fast.
Or be positioned for acquisition. Those are your three paths. Build something truly defensible
through moats we talked about. Generate cash in 12 to 18 months and bank it. Or build something
(21:52):
impressive enough that a giant wants to absorb you. But don't fool yourself into thinking you're
building the next generational super company. Not unless you're in one of those rare categories
with multiple overlapping moats. And even then, stay paranoid.
So next time you see a headline about an AI startup raising $100 million at a billion
dollar valuation, ask yourself, what's their actual moat? Can they get to profitability
(22:17):
before the foundation shifts again? And are they in a category where defensibility is even possible?
Perfect. And remember, 90% of them won't make it. So bet accordingly.
Yaakov Lasker, this has been sobering and fascinating. Thanks for joining us on Innovation Pulse.
Thanks for having me, Donna. Now, if you'll excuse me, I need to go rethink my entire career.
(22:40):
Don't we all? That's it for today's episode. We'll catch you next time when we'll dig into
something that hopefully won't make you question all your life choices. Maybe.
As we wrap up today's podcast, we explored World Lab's new marble platform, transforming
inputs into 3D environments and discussed the challenges AI startups face, with many struggling
(23:06):
due to rapid tech shifts and lacking defensible models. Don't forget to like, subscribe, and share
this episode with your friends and colleagues so they can also stay updated on the latest news
and gain powerful insights. Stay tuned for more updates.
Welcome to Innovation Pulse, your quick, no-nonsense update covering the latest in startups and
(00:09):
entrepreneurship news. First, we will cover the latest news. World Labs unveils Marble,
a platform for creating editable 3D environments, and Cursor raises $2.3 billion to enhance its AI
coding tool. After this, we'll dive deep into the challenges AI startups face, exploring why
(00:32):
many struggle and what makes others succeed in a rapidly evolving tech landscape. World Labs,
founded by AI leader Fei-Fei Li, has launched Marble, an innovative product that transforms text
prompts, photos, videos, 3D layouts, or panoramas into editable and downloadable 3D environments.
(00:55):
Unlike its competitors, Marble offers persistent 3D worlds rather than on-the-fly generation,
enabling users to export worlds as Gaussian, Splats, Meshes, or Videos. It stands out with AI
native editing tools and a hybrid 3D editor that lets users outline spatial structures before
(01:17):
detailing them visually. Marble's flexible input options allow users to upload multiple images
or clips, creating realistic digital twins. Its chisel feature separates structure from style,
offering direct object manipulation, enhancing creative control. The product also allows world
(01:39):
expansion and composer mode for large spaces, making it ideal for gaming, VFX, and VR projects.
Available in four subscription tiers, Marble provides a range of features for various
creative needs. It holds potential for robotics training environments and aims to pioneer spatial
(02:00):
intelligence, a crucial step towards intelligent machines as envisioned by Fei-Fei Li.
Cursor, an AI startup, recently secured a significant $2,300,000,000 funding round,
boosting its valuation to $29,300,000,000. The company offers an AI coding tool designed to
(02:25):
assist developers in generating, editing, and reviewing code efficiently. Originally launched
within Anisphere, a research lab established in 2022, Cursor has quickly become a leading name among
AI startups, valued over $10 billion. The recent funding saw participation from major investors
(02:46):
like Axel, Thrive Capital, Andreessen Horowitz, and others. With this investment, Cursor aims to
deepen its research and enhance its product offerings. The tool has achieved a remarkable
milestone, surpassing $1 billion in annualized revenue and expanding to over 300 employees.
(03:08):
Notably, Nvidia CEO Jensen Huang praised Cursor as his top enterprise AI service.
Despite industry interest, Cursor remains focused on growth rather than an immediate IPO.
The coding tool market is competitive, but Cursor distinguishes itself with proprietary models,
(03:28):
generating more code than most large language models. Competitors include OpenAI and Anthropic,
each developing their AI coding tools, though Cursor continues to lead with its unique capabilities.
And now, pivot our discussion towards the main entrepreneurship topic.
(03:53):
Hey everyone, Donna here on Innovation Pulse, and today I've got someone who's going to make every
AI startup founder either panic or pivot. Yakov Lasker is a senior software development manager
who's been watching the AI startup landscape with what I can only describe as well-informed horror.
Yakov, welcome to the show. Thanks Donna. And yeah, horror is the right word.
(04:16):
You know what keeps me up at night? I just read that 90% of AI startups fail within their first
year. 90%. But here's the kicker, everyone's blaming it on the wrong thing. Okay, you've got
my attention. What's the wrong thing? People keep saying it's because the big tech companies are
more agile than we thought. Right. That OpenAI and Google and Anthropic are just moving faster than
(04:41):
traditional incumbents. But that's missing the entire point. So what is the point? The ground is
literally shifting under everyone's feet every nine to 12 months. It's not that the big players
are outrunning the startups. It's that they're the ones causing the earthquakes that keep knocking
everyone else over. Wait, hold up. Nine to 12 months? That's insane. I mean, most startups need at
(05:06):
least 18 months just to find product market fit, right? Exactly. And that's the trap. Look, I came
across this investment thesis that basically argues every AI application startup is doomed.
And the data's pretty damning. In 2024 alone, ChatGPT went from 200 million weekly users to
(05:28):
400 million by February the 2025. The mobile app generated over a billion dollars in just the first
part of this year. That's 53 times more revenue than its nearest competitor. Okay, but ChatGPT
is ChatGPT. Surely there's room for specialized apps that do specific things better. That's what
everyone thought until OpenAI's Dev Day in October to 2025. They announced they're turning
(05:54):
ChatGPT into an entire platform. Basically an operating system where other developers can build
apps inside ChatGPT itself. Oh, no. Oh, yes. They introduced something called the apps SDK and
Agent Kit. So now instead of users leaving ChatGPT to use your specialized AI tool,
(06:16):
they can just use a version of your tool that lives inside ChatGPT where they already spend all
their time. That's like if Amazon decided to just clone every successful product on their marketplace
and sell it themselves. Oh wait, they kind of do that. Exactly. But it gets worse. Remember all
those AI scheduling assistants. ChatGPT now does automated task scheduling. AI image editors,
(06:43):
ChatGPT has image generation and editing built in. Customer service bots, code completion tools,
search engines. They're just eating everything. Everything. And it's not just OpenAI.
Anthropic went from Claude Sonnais 3.5 in June 2024 to Claude Sonnais 3.7 in February 2025
(07:07):
to Claude Sonnais 4 and Opus 4 by May 2025. That's three major releases in 11 months.
Okay. So I'm hearing a lot of doom and gloom here. But surely some AI startups are making it
work? I mean, we keep hearing about billion dollar valuations. Sure, but here's the thing. Those
valuations are based on potential, not proven business models. There's this fascinating
(07:33):
stat from CB Insights. The 32 new AI unicorns in 2024 represented almost half of all new unicorns,
but they haven't built commercial networks nearly as robust as non-AI unicorns. They're
reaching unicorn status with 203 employees instead of 414 and in two years instead of nine.
(07:58):
So they're growing fast, but not deep. Right. And the money tells you everything. Venture funding
for AI startups dropped 42% in 2024. Most companies that got funded in the 2021 to 2023 boom had 18
to 36 months of runway, which means by late 2025 or early 2026, which is basically now,
(08:23):
they're running out of money. And they can't raise more because investors are getting smarter.
They're getting more skeptical for sure. But here's where it gets really interesting. The
big tech companies aren't even doing traditional acquisitions anymore. They're doing these reverse
acquihires. What's a reverse acquihire? So Google paid $2.7 billion for character AI,
(08:46):
but they didn't buy the company. They just hired the founders and key talent and licensed the
technology. Amazon did the same thing with a robotics AI startup called Cavarian. They took
the founders and 25% of the workforce. Wait, that's brilliant. You get all the talent and the tech
headings of a full acquisition. Exactly. Microsoft did it with inflection. And this is happening at
(09:11):
scale. In 2024, there were 384 AI acquisitions. And a ton of them followed this pattern.
So the two ways to win as an AI startup are what? Build fast, generate cash and get out?
Or build something good enough that Google acquires your team?
That's basically the thesis, yeah. You've got maybe 12 to 18 months to generate cash before
(09:35):
the foundation model providers add your feature to their platform. Or you build something impressive
enough that they want to absorb you. Okay, but I'm hearing a lot of people talk about moats.
Surely there are ways to build defensible AI businesses? There are, but they're very specific.
And this is where the thesis gets more nuanced. The survivors aren't the ones with the best AI
(09:57):
models. They're the ones with the stickiest integrations, the most unique data, or the
deepest vertical specialization. Give me an example. Take a bridge AI. They do medical transcription,
turning doctor patient conversations into clinical notes. But they're not just doing speech to text.
They understand clinical workflows. They've tuned their models on medical terminology.
(10:22):
They integrate with Epic, which is the dominant electronic health record system.
And they embed seamlessly into how doctors actually work. They're now in over 100 health
systems. So it's not about having better AI. It's about being impossible to rip out. Right.
Or take scale AI. They got a five year $100 million contract with the Department of Defense.
(10:48):
To compete with that, you'd need security clearances, SIF infrastructure,
and years of relationship building in government. That's a cornered resource moat. You literally
cannot copy it. Okay, so there's a pattern here. Medical, defense, what else?
Highly regulated industries. Places where you need domain expertise that takes years to build,
(11:09):
or hardware integration, like if you're putting actual devices in people's homes that learn and
optimize over time. The key insight is that the moat can't be the AI model itself, because everyone
has access to the same foundation models. So if you're building on top of Open AI's API,
or Anthropics Clawed, you don't really have a moat. Not a technical one. No. There's this brutal
(11:34):
medium article I saw titled zero moat in AI application startups. The author looked at over
2,000 startups and basically said, if you're just wrapping a nice interface around GPT-4,
you're building on someone else's infrastructure with no defensibility.
And when that someone decides to build the interface themselves, you're done. The article gave this
(11:57):
perfect example. Dozens of AI resume builders launched in 2023, then LinkedIn and Canva just
added the feature natively. All those startups, dead.
Okay, so this is starting to sound less like AI startups are doomed, and more like,
most AI startups are building the wrong thing.
(12:18):
That's a better framing, actually. But the original thesis has this insight that I think is
really important. It's not just about building the wrong thing. It's about the rate of change
being so fast that even if you build the right thing, the foundation shifts before you can mature
as a business. Explain that. Okay, so think about previous technology waves, the PC era,
(12:41):
the internet boom, mobile. Each of those took years to play out. You had time to build a company,
establish sales relationships, develop brand recognition, become an incumbent yourself.
Right. Like how Facebook started in dorm rooms and had time to become the giant that now buys or
crushes new social networks. Exactly. But with AI, the sea changes are happening every nine to
(13:06):
12 months, and very few startups can turn into a mature business in that timeframe. Even if your
engineers can pivot quickly, you can't hire 100 people in a month without your organization
imploding. Sales cycles, customer relationships, brand building, these things are incompressible.
So you never quite make it to the finish line before the race course changes.
(13:29):
That's the metaphor the thesis uses. You can't skate to where the puck is going if 20 people
are about to slap it in unpredictable directions at high velocity. I love that. Okay, but let's get
practical here. If you're an investor or a founder, what do you actually do with this information?
Well, if you're an investor, you're asking very different questions now. Not just how fast can
(13:53):
you grow? But how defensible is this specifically? The thesis suggests looking at whether the
fundamental assumptions underpinning a startup's existence will be the same in five years,
or whether they'll be unpredictably different. And if they're unpredictably different, that's
not investable. Not as a long term independent company. No, you might invest expecting a quick
(14:17):
flip or an acquisition, but you're not betting on the next Google. What about if you're a founder?
If you're a founder, you need to be brutally honest about your moat. Are you building something
that requires years of domain expertise? Do you have proprietary data that gets better with every
user interaction? Are you integrating so deeply into customer workflows that switching to a
(14:39):
competitor would be a nightmare? Do you have regulatory advantages or hardware in the real
world? And if the answer is no to all of those, then you're probably building a feature, not a
company, which might be fine if you can generate cash quickly, or if you're building to be acquired.
But don't fool yourself into thinking you're building the next sales force.
(15:00):
There's another angle here that we haven't talked about. What about open source?
Meta's releasing Lama models for free. Doesn't that change the equation?
It's interesting, right? The data shows that while open source is advancing, there's actually a
consolidation happening around the closed source frontier models. Anthropic now has 32% of the
(15:23):
enterprise market. Open AI has 25% down from 50% two years ago, but Meta's Lama only 9%.
Why is that? I would think free and customizable would win an enterprise.
Two reasons. First, the closed source models are still meaningfully better for most use cases.
(15:44):
Second, and this is crucial, enterprises care about reliability and support.
If something breaks in production and you're using Lama, you're on your own. If you're paying
Anthropic or open AI, you've got a throat to choke, as they say in enterprise sales.
Plus legal liability, compliance, all that fun stuff.
(16:05):
Exactly. Though there is this interesting dynamic where some developers use open source for
experimentation and then switch to closed models for production. But that hasn't translated into
a real business model for the open source model providers. Like, Mistral AI has an unclear path
to revenue according to the information. Less than 10% of users pay for their commercial models.
(16:32):
So open source is great for the ecosystem, but not necessarily for building a business.
Not at the foundation model level anyway. Though some startups are building businesses
by fine tuning open source models for specific verticals. That's a different play.
Okay, I want to circle back to something you mentioned earlier. You said there's a pattern
(16:52):
of successful AI startups in healthcare, defense, regulated industries. Is there anything else that's
working? Yeah, this is where we get into what the thesis calls real atoms data. Basically,
if your AI is dealing with the physical world, hardware, manufacturing, logistics, agriculture,
you might have a shot because that data is much harder to replicate.
(17:16):
Because you can't just scrape it off the internet.
Right. And it often requires sensors, equipment, physical presence. There's a startup called
Goodship that does freight management. They're not just predicting stuff, they're integrated into
the actual operational workflow of moving physical goods around. Replacing them would mean retooling
(17:37):
mission critical logistics infrastructure. So the pattern is go deep, go vertical, go physical,
or go regulated. That's a good summary. And ideally, combine multiple modes. Like that
medical example with average, they've got regulatory modes, workflow integration,
proprietary data that gets better over time, and domain expertise. Any one of those might not be
(18:02):
enough. But all four together, they'd be really hard to kill. Exactly. Though even they're not
completely safe. If open AI decided to partner directly with Epic and built medical transcription
into the EHR. Oh, that's terrifying. Welcome to being an AI application startup in 2025.
(18:22):
Okay, so let's talk about what happens next. We've established that most AI startups are in
trouble, that the foundation keeps shifting, that there are some defensible positions.
What's your prediction for the next 12 to 18 months?
I think we're going to see a massive wave of consolidation.
Bessemer Venture Partners is predicting a surge in M&A activity in 2025 and 2026.
(18:47):
The AI native startups that pushed into specific verticals will get bought by traditional
software companies that need to evolve or die. So like Salesforce buying a bunch of AI startups?
Exactly. Salesforce is already buying Informatica for $8 billion. You'll see more of that.
The traditional SaaS companies realize their products are at risk of being AI-ed out of
(19:10):
existence. So they're buying capabilities rather than building them. Because building from scratch
would take too long. And by the time they finished, the technology would have moved on.
So you're going to see the giants buying the upstarts that actually figured out how to apply AI
in specific domains. The startups get liquidity, the acquirers get to survive.
(19:32):
What happens to everyone else? The ones without moats? They run out of money and shut down.
We're already seeing this. That MIT study showing 95% of business AI implementations failing?
Those aren't just pilots. Those are startups dying because their product got commoditized
before they could scale. And the foundation model providers just keep expanding.
(19:55):
Yeah. They're in this weird position where they're simultaneously infrastructure providers
and application builders. OpenAI is a great example. They provide the API that powers a
thousand startups, but they're also building chat GPT into a platform that competes with all those
startups. That seems like a conflict of interest. It's complicated. From their perspective, they're
(20:19):
just adding features that users want. From the startup perspective, they're pulling the rug out.
And because the foundation model providers control the underlying technology,
they can move faster than anyone building on top of them. So if you're building on OpenAI's API,
OpenAI can always out-execute you because they see what's working and they control the foundation.
(20:42):
In theory, yes. Though I think the reality is messier than that. OpenAI is building for
horizontal use cases. If you're deep in a vertical, like medical imaging or legal discovery,
you might still have an advantage because you understand the problem space better.
But you're always one product announcement away from being obsolete.
(21:03):
That's the existential dread. Yeah.
All right. Let's bring this home. If someone's listening to this and they're thinking about
starting an AI company or investing in one, what's the one thing they need to internalize?
The foundation is not stable. This is not like previous technology waves where you could ride
a trend for five or 10 years. The ground is shifting every nine to 12 months, and it's
(21:26):
shifting in ways you cannot predict with confidence. So whatever you're building,
it needs to be defensible not against competition, but against obsolescence.
And if it's not defensible, it needs to generate cash fast.
Or be positioned for acquisition. Those are your three paths. Build something truly defensible
through moats we talked about. Generate cash in 12 to 18 months and bank it. Or build something
(21:52):
impressive enough that a giant wants to absorb you. But don't fool yourself into thinking you're
building the next generational super company. Not unless you're in one of those rare categories
with multiple overlapping moats. And even then, stay paranoid.
So next time you see a headline about an AI startup raising $100 million at a billion
dollar valuation, ask yourself, what's their actual moat? Can they get to profitability
(22:17):
before the foundation shifts again? And are they in a category where defensibility is even possible?
Perfect. And remember, 90% of them won't make it. So bet accordingly.
Yaakov Lasker, this has been sobering and fascinating. Thanks for joining us on Innovation Pulse.
Thanks for having me, Donna. Now, if you'll excuse me, I need to go rethink my entire career.
(22:40):
Don't we all? That's it for today's episode. We'll catch you next time when we'll dig into
something that hopefully won't make you question all your life choices. Maybe.
As we wrap up today's podcast, we explored World Lab's new marble platform, transforming
inputs into 3D environments and discussed the challenges AI startups face, with many struggling
(23:06):
due to rapid tech shifts and lacking defensible models. Don't forget to like, subscribe, and share
this episode with your friends and colleagues so they can also stay updated on the latest news
and gain powerful insights. Stay tuned for more updates.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.