Speaker A
If you are building a startup on top of an OpenAI or Anthropic API, there's a real chance that your startup is already dead and you don't even know it. I'm not saying that because your product is bad,
Startups built on OpenAI APIs risk being replaced as labs integrate their features; success depends on owning unique data, workflows, or fine-tuned models.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- Building solely on top of AI model APIs is risky as labs can integrate your features and replace your product.
- Success requires owning unique data, workflows, or fine-tuned models that labs cannot easily copy.
- Fine-tuning open-source models on narrow tasks is a cost-effective way to maintain control and differentiation.
- The AI value chain favors hardware providers and application layer companies, while model labs face margin pressure.
- Startups should treat API wrappers as temporary and focus on building defensible assets for long-term viability.
What the video covers
- Startups relying solely on OpenAI or Anthropic APIs risk obsolescence as these labs integrate popular features directly into their platforms.
- OpenAI and similar labs act like suppliers who become competitors by embedding customer innovations as native features.
- The AI ecosystem's value pools at the compute hardware level and the application layer, while model labs face tough economics and price pressure.
- Labs can cheaply replicate successful use cases because they pay wholesale for models and control distribution to hundreds of millions of users.
- Startups must evaluate if improvements in AI models make their products more or less valuable to determine their sustainability.
- Three defensible assets that labs cannot easily replicate are exclusive data, deeply integrated workflows, and fine-tuned models on proprietary tasks.
- Fine-tuning smaller open-source models on narrow tasks can reduce costs, increase control, and protect data privacy from labs.
- Using API wrappers is a valid initial strategy but should be a stepping stone towards building unique assets that labs cannot copy.
- Startups must build versions of their products that remain viable even if OpenAI or similar labs disappear, ensuring independence.
- The key to surviving and thriving is owning the 'last mile' of customer value that labs cannot commoditize or replicate.
Chapters
- 00:00The Risk of Building Startups on OpenAI APIs
- 00:20Suppliers Becoming Competitors in AI
- 00:38The AI Ecosystem and Money Flow
- 01:32Model Labs' Economics and Competition
- 02:24The Impact of Custom GPTs and Feature Integration
- 02:52Why Labs Can Copy Your Product Easily
- 03:16Key Question: Does a Better Model Help or Hurt Your Product?
- 04:05Three Assets Labs Cannot Replicate
- 05:18Fine-Tuning Models for Narrow Tasks
- 06:07Building Defensible Startups Beyond Wrappers
Full Transcript — Download SRT & Markdown
Speaker A
but because the company that sells you the intelligence has quietly taken note of your best idea and potentially decided that it's their next feature.
Speaker A
Your product will become their feature. A VC put the entire AI ecosystem very aptly. OpenAI built a giant farm that grows tokens and sells them to everyone.
Speaker A
Then it noticed that money wasn't in growing the crop. It was in the restaurants people were building with it. So the company started opening its own restaurants right there on the farm.
Speaker A
And every founder who was buying vegetables looked up and realized that their supplier just became their competitor. That's the fear going around right now. And yet some startups in that exact same position are going to come out on the other side worth billions.
Speaker A
The difference between the ones that die and the ones that win isn't luck. And it isn't the model that they choose. It's one thing and it's learnable. And by the end of this video, you will know which side of that line you are on and how to
Speaker A
get on the right one. Something strange happens when you follow the money up the AI stack. At the bottom is Nvidia selling the shovels, keeping something like 70 cents on the dollar. $75 billion a quarter from data centers alone. At
Speaker A
the top is the application layer. Everyone is building products. And in the middle sets the model labs, the supposed kings of this entire ecosystem with the worst economics of anyone involved. OpenAI is doing around $20 billion a year and losing around 14 of
Speaker A
that and their product raw intelligence getting cheaper every quarter as three rival labs and a pile of open-weight models drag the price lower and lower.
Speaker A
That's the current reality. Value running out of the middle and pooling at the two ends. The compute below which stays scarce and the last mile above where the customer is. As we speak, the labs' margins are ticking up because
Speaker A
everyone wants agents and compute is scarce. So they can charge more per token. Think about electricity. When power became cheap and universal, the value didn't just disappear. It moved to everything that you plug into. Nobody builds a great company today by
Speaker A
reselling electricity. The model is becoming that electricity. Reselling it with a nicer login screen is not really a business anymore. Jasper, an AI writing startup, was worth 1.7 billion, right up until the thing it wrapped became free inside ChatGPT. At OpenAI's
Speaker A
first developer day, one announcement, custom GPTs wiped out the whole category of small startups overnight. A room full of founders watched their product turn into a checkbox. And coding tools like Cursor now watch OpenAI's Codex and Anthropic Squad Code walk straight onto
Speaker A
their turf. The suppliers shipping the exact product that its customers built. Now, none of this even needs ill intentions. The lab simply sees which use cases are taking off and because it can watch demand right through the API,
Speaker A
it can copy them. It owns the cheaper path because it pays wholesale for the model which you pay retail for and it owns the distribution. It can put a feature in front of 800 million people on a random Tuesday. What Alex Karp
Speaker A
actually meant was a little more honest. His line was, "If the value is so obvious, why are they selling you the tokens at all? If I could make you a billion dollars, wouldn't I take a cut instead of selling you the raw
Speaker A
ingredients?" The labs aren't breaking in, they're climbing up the stack towards you because that's where the money is, which is arguably worse because it's completely rational. If you are wondering which side of the fence you are on, try answering this one
Speaker A
question. If the models got twice as good tomorrow, smarter, cheaper, or faster, does your product get more valuable or less? If the answer is less, if a better model makes you more replaceable, you have built a wrapper and the wrapper will get wrapped. If a
Speaker A
better model makes you more valuable because it plugs into something the lab doesn't have, you have built a company.
Speaker A
Could a sharper team clone your whole product in a weekend? If you can't name a single asset a lab couldn't copy in a quarter, that's not an insult. That's your to-do list. There are three assets a lab can't ship as a feature. The first
Speaker A
is data it cannot see. Open Evidence, a medical answer engine, looks like a ChatGPT clone, except it runs on an exclusive license to the New England Journal of Medicine, and it's used by two-thirds of American doctors. A lab can
Speaker A
copy that inference in a weekend. It cannot copy the license or the trust. That's the moat. A data flywheel that gets better every time a customer uses it. Built on data that's never touched by the labs. Someone can copy your
Speaker A
features in a quarter. They cannot copy three years of your customers' private data. The second asset is workflow. Because a system of record wired so deeply into how the customer operates that pulling you out breaks their week. That's
Speaker A
Palantir's whole game. Not a smarter model, but 20 years of integrations, audit trails, and plumbing that nobody really wants to rebuild. The third is that you don't have to beat the frontier at everything. You just have to beat it
Speaker A
at your one job. Take an open source model like LLaMA, Qwen, or Mistral and fine-tune it on your task, on your data.
Speaker A
There's a landmark result from a company called Preybase. Twenty-five little 7 billion parameter models each fine-tuned on a narrow task matched or beat GPT-4 on those exact things for under $8 of training apiece. And when you own that
Speaker A
model, three things change at once. Your per token bill falls off a cliff. Nobody can silently swap the model out from under you. And your data and your usage live somewhere the lab will never see.
Speaker A
You commoditize the model before it commoditizes you. Fine-tuning a small model wins on narrow well-defined jobs.
Speaker A
Classification, extraction, or specific structured output. Point it at an open-ended reasoning or real coding and it falls apart. And the second you own the weights, you own a small infrastructure company. GPUs and ML engineer salary and a retraining
Speaker A
treadmill to fight the models going stale. So what do you actually do? Well, firstly, use the wrapper. The fastest way is still to wrap an API and get something real in front of users right now. Just treat that wrapper as a
Speaker A
stepping stone, not the entire solution. Every week it's alive. Spend the runway acquiring the one thing the lab can't ship. A data loop only you have. A workflow they can't tear out or a fine-tuned model that they can't see
Speaker A
or swap. The intelligence is for rent now to you and to everyone else competing with you. What you get to own is the last mile. Stop renting your whole company from your competitor.
Speaker A
Build the version that still stands if OpenAI vanishes tomorrow. That's the version they can never ship as a feature and it's the only line between startups that die and the one that everyone else ends up copying.
Topics:OpenAIAI startupsAnthropicAPI risksAI ecosystemfine-tuning modelsdata moatworkflow integrationAI business strategymodel labs economics
