What Startup Spending Data Reveals About AI Business Models
- Baran Korkut
- Oct 4
- 7 min read
When I wrote about how AI works in business design, my main argument was that AI didn't fundamentally change business design principles, it just made things easier and faster. Now, we have concrete evidence of what's actually happening in the market. And it turns out the fundamentals still hold, but the game has changed in ways that matter.
Andreessen Horowitz recently published spending data from over 200,000 startups, showing where real money flows in AI applications. This isn't web traffic or hype metrics. This is actual purchases, what founders pay for when they're building companies. The patterns in this data tell us something important about how business models are evolving in the AI era.
The Fundamentals Haven't Changed (But Everything Else Has)
The a16z data confirms something I've advocated for long: AI is an accelerant, not a paradigm shift. Companies that make the top 50 spending list—OpenAI, Anthropic, Replit, Notion, Canva—are doing what winners did in the pre-AI era. They're creating the best value for their focused customer segments.
What's interesting is that Notion and Canva made this list at all. They're not AI-native companies. They're established products that integrated AI effectively. This suggests that a solid strategy to incorporate AI into existing products does exist. The key is still adding value to customers, whether you use AI or not. You can add more value with AI, but you still need to obsess about the customer problem.
But while the principles remain constant, several structural shifts are reshaping how business models actually work.
If You Can Dream It, You Can Build It (And So Can Everyone Else)
AI capabilities are becoming commoditized. Everyone has access to similar models. But there's a more significant phenomenon at play: if you can dream it, you can build it with AI.
We know ideas are cheap and plenty. Implementation is what matters. With AI, you can implement. Look at the adoption of vibe coding tools at the enterprise level. Replit, the top product on the spending list after OpenAI and Anthropic, generated roughly 15x more revenue than Lovable from the Mercury customers analyzed. These tools aren't just helping developers work faster, they're enabling non-developers to build functional applications.
When implementation barriers drop, companies can build more complete products. Not necessarily more innovative products, but ones that tick more boxes, more workflows, more features, more coverage of the problem space. This completeness is what catches enterprise attention.
But here's the catch: if you can build it, so can your competitors. And so can your potential customers.
Enterprises Are Getting Smarter (And Pickier)
The same implementation barriers dropping for startups are also dropping for enterprises. Corporate innovation teams can now spin up their own AI-powered prototypes to test ideas internally. This doesn't give them the operational agility to compete with startups, but it does make them smarter buyers.
When enterprises can run their own experiments, they develop more complete asks from startups. They understand the problem space better. They know what's possible. This might sound like a threat to startups, but it's actually an advantage. More complete requirements make it easier for focused startups to perform.
The traditional bottleneck in business design has always been commercial risk, will people buy this?, not technology risk. Commercial risk remains higher. But now enterprises can validate some of their assumptions themselves before buying, which means when they do buy, they're more committed to solving the problem fully.
The Return of Pure Product-Led Growth in AI Business Models
Here's what the spending data shows: 70% of companies on the top 50 list can be adopted by individuals and brought into teams without requiring an enterprise license. Eleven of these companies started as primarily individual products and evolved to offer team or enterprise functionality over time.
Several are still generating consumer-majority revenue. OpenAI went from 75% consumer revenue to roughly a 50/50 consumer-enterprise split in less than a year. This isn't just faster PLG. It's a structural shift.
In previous software generations, you had to design for two customer segments: the individual as the user and the corporation as the customer, because the company pays. Now, for this subset of AI startups at least, founders only need to obsess about the user. The company follows.
The product spreads because individual employees use it for their work. Eventually the company picks up the tab. This consumer-to-enterprise flywheel is happening in 1-2 years instead of the multi-year journeys we saw before AI.
Maybe we're full-circling back to pure PLG. The product is great, so it sells. The difference is that these products' customers are also corporate employees using the products for their corporate jobs.
When Everyone Can Code and Design (Sort Of)
The a16z data identifies creative tools and vibe coding as categories that have become "horizontal", used by people in any role, not just specialists. This is one of the more interesting patterns in the data.
But let's be clear about what's happening. These tools don't make everyone a standout designer or quality developer. They allow anyone to tap into a general intelligence that can put together shapes and colors or pull bits from GitHub. What they really enable is thinking with artifacts, externalizing ideas quickly.
Consider it this way: there are good coders and non-coders. Now both have a new capability called vibe coding. The difference between them is still the difference between a good coder and a non-coder. But both gained a new tool.
Take Lovable, which ranks high on enterprise purchases. It's definitely a non-developer tool. Corporations aren't buying Lovable for their development teams. They're buying it for engineering, marketing, operations, people who can use code to do their jobs better but don't want to wait for a developer to save them.
The message is about removing dependencies and bottlenecks within organizations, not replacing specialists. If you can use code to improve your work, use code. Don't wait.
This horizontalization expands the market for these tools. They serve both non-experts who can now produce basic work themselves and experts who can work significantly faster. Lovable leans toward non-experts. Cursor is beloved by experts. But both sell to both segments.
Augmentation Wins Over Replacement
Of the 17 vertical AI tools in the top 50, twelve focus on augmenting humans in their roles. Only five aim to be full "AI employees" completing workflows end-to-end. I don't think replacement will ever win over augmentation for complex work. For data entry, sure. But for knowledge work, the dynamic is different.
Assume AI has some level of intelligence and so does a person. We don't need to argue who's more intelligent. What matters is that there's true synergy in human plus AI. AI performs better with capable human guidance. Humans perform better with quality AI assistance.
This isn't a transitional phase until AI gets better. The synergy thesis seems fundamental. Human plus AI is better than either alone for knowledge work, and this will likely remain true.
Community as Distribution
One pattern that doesn't show up explicitly in the a16z data but seems relevant: AI startups appear to have strong contributing communities. Users help each other, create and share resources, and evangelize the product.
This might be becoming a required component of the business model. When individual users are both your target customer and your distribution channel into enterprise, community becomes critical for turning users into product advocates.
From a design perspective, founders just need to point to a meeting place for the community. If you create value for customers, they'll form a community if they have a place to go. This can be as simple as a subreddit.
The community isn't just support infrastructure. It's part of the customer relationship strategy on the business model canvas. It's how users become advocates who pull your product into their organizations.
What Should Founders Do Differently?
Every founder building in this environment needs to find a way to make a personalized general LLM assistant work effectively with them. This is no longer a nice-to-have or an outlier approach. AI is your first hire. Then you can move onto more specialized AI tools if necessary.
This might sound like obvious advice, but it's tactical. You need to build your own human-plus-AI working rhythm before you can effectively design AI products or integrate AI into your business model.
Beyond that, the advice remains surprisingly traditional. Create value for focused customer segments. Don't skip customer discovery and validation just because you can build faster. AI makes the build phase faster and more complete once you've validated, but it doesn't excuse you from understanding what customers actually need.
If there's one thing that deserves new emphasis given what the spending data reveals, it's this: it makes more sense than ever to add "with AI" to the end of the most crucial question in business design, how can we solve this big problem?
Not "how can AI solve this problem" but "how can we solve this problem with AI as a tool."
The difference matters.
The Gap Between Hype and Adoption
What strikes me most about the a16z spending patterns is what they reveal about where companies are actually finding value versus where the hype suggested they would. The top of the list isn't dominated by autonomous agents or revolutionary new categories. It's dominated by tools that make existing work faster and easier. Meeting notetakers. Creative tools that produce decent outputs quickly. Coding assistants. General LLM interfaces.
The winning business models aren't the most technically impressive. They're the ones solving real problems for real users in ways that integrate into existing workflows. Replit generates 15x more revenue than Lovable in enterprise not because it's more AI-native, but because it provides enterprise-grade functionality—databases, authentication, secure publishing—alongside the AI capabilities.
The lesson for founders is the same lesson that's always been true: focus on the customer problem, not the technology capability. AI gives you more tools to solve problems, but it doesn't change the fact that you need to solve actual problems.
The Principles Still Hold
AI didn't fundamentally change business design principles and process. It made things easier and faster. We see the same effect throughout the spending data. AI has made horizontal and vertical tasks easier and faster across categories.
But easier and faster compounds. When PLG cycles compress from years to months, when implementation barriers drop for both startups and enterprises, when creative and coding work becomes accessible to non-specialists, when communities can form and scale quickly, these accelerations change the competitive landscape even if the underlying principles remain constant.
The startups winning in this environment are the ones that understand both halves of the equation. The fundamentals still matter. Customer obsession still matters. But the game moves faster now, and the business models that succeed are the ones designed with that speed in mind.
AI is here to help. The question is whether you're using it to solve real problems or just to build faster solutions to problems that don't exist.
Comments