In 2023, the business model was clear.
Anthropic builds models. Developers access them via API.
Developers build products. Users use the products.
Anthropic makes money from the API. Everyone wins.
This was the foundation on which thousands of companies built
their products, raised their funding, and hired their teams.
The labs were infrastructure. The ecosystem was the product layer.
The relationship was symbiotic and straightforward.
That model is gone.
It did not disappear with an announcement. It dissolved gradually,
product launch by product launch, feature addition by feature addition,
until the labs that were powering the ecosystem were also competing
with most of it.
The developers still paying for API access are funding the
products being built directly against them.
What the pivot actually looks like
Claude.ai is not a research demo. It is a full-featured consumer
and enterprise product. Memory. Projects. Artifacts. Voice.
Deep research. Integrations. The surface area of what it does
has expanded past what most Claude-powered startups offer.
ChatGPT is the same. The product that used to be a simple
chat interface now includes image generation, code execution,
web browsing, file analysis, memory, custom GPTs, and a canvas
for collaborative document creation.
Google's Gemini is embedded into every enterprise product Google
sells. Docs. Sheets. Gmail. Meet. The AI is not an API endpoint
anymore. It is a feature of the productivity suite two billion
people use at work.
Each of these is a direct consumer and enterprise product.
Each competes with companies built on the APIs that fund their development.
A legal AI startup built on Anthropic's API is now competing
with Claude.ai's document analysis features. A writing assistant
built on OpenAI's API is competing with ChatGPT's canvas.
An enterprise search product built on Google's API is competing
with the AI search built into Google Workspace.
The infrastructure became the application.
The platform playbook everyone forgot
This is not the first time a platform company has done this.
Microsoft built Windows and charged developers to build on it.
Then Microsoft built Office to compete with the most successful
apps built on Windows. The antitrust case that followed changed
how the industry thinks about platform power.
Apple built iOS and took a thirty percent cut of everything
sold through the App Store. Then Apple built its own apps
that competed with the most successful App Store categories.
The developers who built podcast apps, navigation apps, and
cloud storage apps watched Apple ship native versions of each.
Amazon built AWS and charged companies to run their infrastructure
on it. Then Amazon used the data from those companies to build
competing products. Several lawsuits and congressional hearings
later, the practice is documented but continues.
The AI labs are following an identical playbook.
Build the platform. Charge the ecosystem to build on it.
Observe which use cases are most valuable. Build native versions
of those use cases. Compete directly with the developers whose
API spend funded the capability that now threatens them.
The difference this time is the speed.
Microsoft took decades to move from platform to application competitor.
Apple took years. The AI labs are moving in months.
The funding that powered the competition
Here is the specific dynamic that makes this more uncomfortable
than the historical analogies.
The companies that raised venture capital to build AI products
on top of these APIs were not just paying for inference.
They were funding the labs' research, their scaling, their
product development, their hiring.
Every dollar spent on the Anthropic API in 2023 and 2024
contributed to building Claude.ai. Every dollar spent on the
OpenAI API contributed to building the ChatGPT features now
competing with OpenAI's developer ecosystem.
The developers were, in a meaningful sense, paying to build
their own competition.
This is not a conspiracy. The labs needed revenue to develop
the models. The developers wanted access to the models.
The transaction was transparent.
What was not transparent was the destination.
The labs were not building API businesses. They were building
AI companies that happened to have API revenue as an early
business model. The API was the revenue mechanism that funded
the product they actually wanted to build.
The developers who treated API access as a durable competitive
position were misreading what the labs were building.
The enterprise deal that changed the dynamic
The consumer product competition is visible and discussed.
The enterprise dynamic is less visible and more consequential.
The major AI labs have all signed significant enterprise contracts
directly. Not API access contracts. Full enterprise AI platform
contracts. The kind that include deployment, support, compliance,
integration, and training.
These contracts compete directly with the enterprise AI companies
that built on top of the same APIs.
A startup that spent two years building an enterprise AI platform
on Anthropic's API is now being asked by prospects to explain
why they should pay for the startup's platform rather than
buying Anthropic's enterprise offering directly.
The answer used to be easy. The startup had features Anthropic
did not offer. The startup had integrations Anthropic had not
built. The startup had enterprise support Anthropic was not
providing.
Each of those advantages is shrinking as the labs invest in
direct enterprise sales.
The startups are not losing because their products are bad.
They are losing because the company that sold them their raw
material decided to sell the finished product instead.
What the developer relations team says
Ask the developer relations teams at any major AI lab whether
they see the tension in this.
The answer is polished and consistent.
The API will always be available. The developers building on
it create use cases the labs could never build themselves.
The ecosystem makes the platform more valuable. We are not
competing with our developers. We are expanding the market.
This is the same answer Microsoft gave in the 1990s.
The same answer Apple gives to this day.
The same answer Amazon's AWS teams give when asked about
Amazon using seller data to build competing products.
Platform companies are constitutionally incapable of
acknowledging the conflict of interest that is inherent
in their position. Acknowledging it would undermine the
developer trust that makes the platform valuable.
So they do not acknowledge it.
And the developers, who need the platform access and have
no real alternative, accept the answer because the alternative
is not building at all.
The products that survive this
The companies built on AI APIs that are not being disrupted
by lab-direct competition share a specific characteristic.
They are not doing something the lab cares about doing itself.
The niche is too vertical. The customer is too specialised.
The workflow integration is too deep into a specific industry
that the labs do not have the domain expertise to serve directly.
A general-purpose writing assistant is directly in the labs'
crosshairs. A writing assistant specifically for clinical
trial documentation that integrates with FDA submission
workflows and maintains 21 CFR Part 11 compliance is not
something Anthropic is building.
The horizontal products are being squeezed. The vertical
products have more runway.
This is the same advice that survived every previous platform
transition. Go vertical. Go deep. Go where the platform
company cannot follow without your domain expertise.
The advice is not new. The urgency is.
The VC money that knew
Here is the conversation that happened in many venture capital
partnership meetings in 2023 and 2024 that the founders
pitching in those meetings did not hear.
One partner: this is a good product but they are building
directly on OpenAI's API. What happens when OpenAI ships this?
Another partner: they have a six to twelve month head start.
If they can build network effects or proprietary data before
that, they survive. If not, they get acqui-hired.
The first partner: should we tell them?
The second partner: they know. Everyone knows.
The founders knew. The investors knew. The outcome was understood
as a risk that could be beaten by moving fast enough to build
something defensible before the lab shipped the same thing.
Some teams beat it. They moved fast enough, went deep enough,
built enough proprietary data or network effects that they have
a position the lab cannot easily replicate.
Most did not move fast enough.
The lab shipped.
The question worth asking now
For every team currently building on AI APIs, one question
determines the next two years.
Is what we are building something the labs will want to build
themselves?
If the answer is yes, the timeline to competition is not
a question of whether but when. The features the labs ship
internally start as limited rollouts to enterprise customers.
Then they expand. Then they become part of the standard product.
The third phase is when the API-dependent company discovers
that the market it was building for has been absorbed.
If the answer is no, the question is why not. The answer
should be specific. The customer is too vertical. The workflow
integration is too specialised. The regulatory requirement
is too complex. The proprietary data advantage is too durable.
Vague answers are not reassuring. The labs' product scope
has expanded faster than anyone predicted in 2023.
What is out of scope today may be core product in eighteen months.
The developers who thought they were building on infrastructure
were building on a foundation that had a different future in mind.
That future is arriving faster than the runway of most companies
built on it.
The API is still available.
The question is whether the thing you built with it
is still defensible now that the company selling you
the API is also selling the product.
For too many teams, the honest answer is no.
And the time to have asked the question was two years ago.
Top comments (0)