🔥When Building Gets Easy, Winning Gets Hard
How winning AI companies are built
👋 Hi, I’m Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with Data & AI. Join 34,870 thought leaders from VCs like a16z, Accel, Index, Sequoia, and more to understand how startup investing becomes more data-driven, why it matters, and what it means for you.
ICYMI, check out our most read episodes:
Brought to you by Affinity - The State of AI Investing
AI is moving from concept to reality in private capital. Hear from experts at BlackRock, OpenAI, and Affinity as they explore how AI is reshaping workflows, unlocking new insights, and redefining what’s possible for deal teams.
This session is built for investors looking to understand where AI can drive real value and how to assess readiness across their portfolio and the broader ecosystem.
Entering the fast fashion era of SaaS..
I shared my take on Sam Altman’s recent tweet that “we’re entering the fast fashion era of SaaS very soon” on LinkedIn here.
The post triggered a controversial discussion with 139 comments (and counting) - something I really enjoy and one of the key reasons I regularly post online. Few better ways to challenge your thoughts at scale ;)
Today, I’d like to dive deeper on this topic and share:
How software development has changed and why there’s only 2 categories left: the hard and the easy build
“When building remains hard, winning gets easy. When building gets easy, winning gets hard.” - and how you can still win in an era where everyone can build
Scale, moats, and margins across the stack: Why the hard build gets rewarded today but the easy build has hope for tomorrow
Let’s dive in!
There’s only 2 categories of software left: the hard and the easy build
There’s no software company without AI anymore. This was different not too long ago.
Deconstructing the AI Landscape (2020)
In 2019, I conducted an internal deep dive at Earlybird about the implications of AI on businesses and established a simple framework that I also published in my “Deconstructing the AI Landscape” article.
Do you want to build a core-AI tech component which is applicable to variety of use cases and independent from a specific industry (horizontal), or do you aim to solve a very use-case- or domain-specific issue by applying AI (vertical)?
With this simple question, I split AI businesses into horizontal and vertical ones, and explored what it takes for either one of them to win.
Re-reading my medium post today, I’m frankly a bit surprised how little my perspective on the fundamentals has changed.
Yes, AI dominates the news and we keep pushing the boundaries on a daily basis. And of course we reflect this in more detailed frameworks as the one below from my 2023 “AI Cheat Sheet”.
But in its core, if we simplify to the lowest denominator, we only have 2 categories of AI businesses.
The two categories of AI businesses in 2025
While in the past I called them “horizontal / core AI / industry agnostic” and “vertical / application / industry specific” companies, I would slightly reframe and simplify even further today:
Hard to build - aka the horizontal ones: These are the companies that despite transformative AI innovations remain notoriously hard to build. They require deep technical expertise and sophisticated software development skills to create and scale complex architectures.
Examples include infra companies (like Nvidia, HPE, Intel, GCP, Azure, AWS - and our portfolio company Arago), intelligence & foundation model companies (like Black Forest Labs, ElevenLabs, Mistral - and our portfolio company spAItial), middle layer companies (i.e. the picks and shovels like our portfolio companies TopK for hybrid databases or Aiven’s open source data platform), but also full-stack platforms (like Antrophic, OpenAI, Databricks, Microsoft, Google - and our portfolio company Aleph Alpha).Do they use AI? Of course. Does it accelerate their delivery? Surely. And does AI reduce the entry barriers to build better competitors with ease? Well, actually no. At least not today. All of the above companies remain hard to build. AI or not.
Easy to build - aka the vertical ones: Building these companies has massively changed in the last 12 months. With Lovable, Bolt, Cursor, Replit and likes we’ve seen entry barriers to build vertical or use case specific solutions almost disappear. Examples include anything you can classify as an “AI wrapper” such as the code-gen tools mentioned above but also (rather simple) use case specific products like DocuSign.
What it takes to win
“When building remains hard, winning gets (relatively) easy.”
For the first category of companies that remain hard to build, I’d say that once the build has been completed and the metrics indicate PMF, the easier part is to scale the customer acquisition and win.
Sounds silly? Yes, I know. No win is easy.
But for these companies, the harder part is definitely finding the right talent - the AI infra engineer or the LLM researcher, the kind of people that receive these $MM offers from Meta these days - and pulling off this hyper complex product.
Once this is done, the scale becomes easier as there are very few alternatives in the market that have solved the same problem and shipped a great product, that it’s relatively easy to stand out and acquire customers. There’s just less noise.
“When building gets easy, winning gets (exponentially) hard.”
As entry barriers disappear, it becomes easier to build AI application layer companies, and as a result, the number of similar solutions starts exploding.
Speed over perfection. Products will be launched faster than ever. In line with Sam’s “fast fashion” analogy, features will be trendy, disposable, and iterated at breakneck speed. Once the product v1 is built and demand got proven, the hard part begins:
➡️ Go hyper niche or go broke. Buying behaviours seem to split into two camps: hyper-niche best-of-breed versus all-in-one platforms. You don’t want to end up somewhere in-between. While all-in-one platforms tend to fall into the camp of “hard builds”, the “easy builds” are likely the be hyper-niche products. Don’t go to broad too early and rather become the go-to solution for one very specific problem.
➡️ Building a solid v2. You don’t need an engineer to build the v1. But once demand got proven and your early user base keeps growing, you should double down and solidify the product sooner than later. You hire a few engineers. Not average ones, but 10x engineers. Reducing churn and improving retention will be key, as eventually this will drive your customer LTV.
➡️ Acquiring customers at low CACs. Standing out in a flood of look-a-likes is hard, and will become harder every single day. Therefore, it becomes critical to build content & community early, no matter if via founder personal brand, your own newsletter or podcast, events, Slack channels, or other means.
In summary, go narrow and deep, and find an unfair advantage for distribution as early as possible.
5 phases to win the AI app market
Phase 1 “The Weekend Wonder”: v1 got vibe-coded over the weekend. Customers sign up. They love it. Traction explodes. Churn too. How to solve it?
Phase 2 “Building the Foundation”: That’s where many companies raise external capital based on their initial momentum with the clear goal to hire some 10x engineers and solidify the product. If done right, churn decreases, retention improves, LTV grows.
Phase 3 “Distribution Crunch”: If you’ve baked in distribution from day 1, your CACs remain healthy as they scale. If not - and that’s the reality for majority of AI application founders - you’ll find yourself throwing more money on paid ads, soon seeing CACs sky-rocket. That’s the latest point in time where you should hire a growth expert and test novel distribution channels, as described above.
Phase 4 “Power-User Paradox”: LTV/CAC ratio becomes sustainable. Number of users and usage per user keeps growing. You identify what power users look like but in return also notice that you actually pay extra for their heavy use, driving your LLM costs through the roof. Margins? Ah yes, totally forgot about this.
Phase 5 “Margin Mastery”: In order to improve margins, you have two levers: The top line and the bottom line. For the top line, that’s where most companies revise their pricing. Usage based, outcome based, credits, tokens - you name it. For the bottom line, that’s where for the first time you not only look for LLM capabilities but for cost per token too. Intent classification to find the most suitable model and not always just use the best that’s out there? Yes please. Moreover, you should get in touch with model providers to negotiate volume based partnerships.
Searching for margins
In light of recent articles by TechCrunch and The Information, and strongly related to the subject of this post, I briefly want to share my perspective on the margins of AI application startups.
Even if you find a great niche (phase 1), solve the churn issue and grow LTC (phase 2), win the attention battle and reduce CACs (phase 3), and identify the power users (phase 4), the most debated challenge these days is phase 5 and whether AI application startups will ever be able to achieve sustainable margins.
Across the AI stack, the picture looks roughly like this:
Application layer: often low or even negative gross margins, as just reported for Replit, the AI coding assistant. According to The Information, its gross margins have been fluctuating between –14% and +36%. For context, in the golden age of SaaS, anything below 60% was considered weak, and best-in-class businesses regularly hit 80% or more. Today, many AI application companies are operating in a very different reality.
Intelligence & foundation layer: ~50% gross margin, ranging from OpenAI at 40-45%-ish to Antrophic at the 55-60%-ish; take it with a grain of salt because private companies and no definite info available here
Infra layer: ~70% gross margin, ranging from TSMC & ASML at the 55-60% range to Nvidia at 75%
Hovering around zero at the top, somewhere around 50% at the middle, and around 75% at the bottom. That’s true gravity.
The Bull case for margin recovery at the application layer
Being an AI optimist, I believe these thin margins are likely temporary. Here are some core beliefs supporting a margin improvement on the application layer:
Falling inference costs: As hardware improves and models become more efficient, the cost of each token processed will decline.
Token-free features: Products can shift some functionality away from the model and onto cheaper, deterministic systems once workflows are well-understood.
Model switching: Many use cases don’t require the absolute state-of-the-art (SOTA) model. Switching to smaller, cheaper models for certain tasks can dramatically reduce costs without sacrificing user experience. That’s the router logic described in phase 5 above.
If these trends play out, gross margins at the application layer could eventually reach the healthy levels SaaS investors are accustomed to. Of course, I’m well aware of the anti-thesis “race to the bottom” perspective but I do believe the upside case is more likely. Call me a structural optimist ;)
Stay driven,
Andre
Thank you for reading this episode. If you enjoyed it, leave a like or comment, and share it with your friends. If you don’t like it, you can always update your preferences to reduce the frequency and focus on your preferred DDVC formats. Join our community below and follow me on LinkedIn or Twitter to never miss data-driven VC updates again.