Daniel, founder of a new AI startup, recently scaled his AI-powered SaaS app to $250,000 in annual revenue. It happened fast, and he was thrilled. The product was taking off, users were growing, and everything looked like it was working. Then came the shocker: a cloud invoice for $800,000, driven almost entirely by inference and compute tied to API usage.
The company had grown the top line, but not the margin. It was scaling itself out of business.
This kind of story is becoming more common as we move into the AI era. The old SaaS playbook of build a great app, charge monthly and let infrastructure fade into the background, doesn’t hold up when your core cost scales with usage.
AI has reshuffled the value chain, and for startups, this shift is existential.
The AI stack is deep and margin has moved
In traditional SaaS, most of the value was captured at the application layer. Today, AI companies operate in a much deeper stack:
Energy infrastructure: Data centers, cooling and power (see Amazon’s $10 billion investment in data center energy in Virginia);
Chips and hardware: Nvidia’s H100s, Google TPUs, scarce and expensive;
Cloud platforms: Azure, AWS, GCP with priority GPU access;
Models: OpenAI, Anthropic and increasingly open-source players;
Vertical AI solutions: Can be used as low code/no code platforms to build specific AI applications; and
Applications: The user-facing product, where most AI startups still live.
But unlike the past, margins no longer concentrate at the top, close to the end user. They now often sit below the surface, especially in layers where scarcity exists such as hardware, compute and exclusive model access.
So what can startups do when they don’t own the infrastructure or the models?
Three moves founders can make to stay in the game
1. Own your data. It’s your new moat
You don’t need to train your own foundation model, but you do need to own the inputs that make your product valuable.
If you’re in a vertical such as healthcare, finance, real estate or legal, your advantage is proprietary, structured data. Fine-tune open models. Build lightweight adapters. Use your customer workflows to continuously collect differentiated data. The value is in the dataset.
2. Price for usage, not access
That founder’s $800,000 cloud bill happened because they were charging like a SaaS company but operating like a compute company.
In AI, usage drives cost. That means flat-rate subscriptions don’t work. Founders must embrace pricing models that align value delivered with cost incurred:
Per-output or per-token billing;
Compute-aware pricing tiers; and
Charging for high-cost features such as image generation or live inference.
Track gross margin by feature, not just customer.
3. Avoid model lock-in. Design for flexibility
Tying your roadmap to one model provider like OpenAI or Anthropic is risky. Latency, pricing and policy changes can all blindside you.
Instead, build with model abstraction in mind. Route across providers, fine-tune open-source backups, and negotiate contracts with leverage. Flexibility is not just technical. It is a business hedge.
Itay Sagie is a strategic adviser to tech companies and investors, specializing in strategy, growth and M&A, a guest contributor to Crunchbase News, and a seasoned lecturer. Learn more about his advisory services, lectures and courses at SagieCapital.com. Connect with him on LinkedIn for further insights and discussions.
Illustration: Dom Guzman
Stay up to date with recent funding rounds, acquisitions, and more with the
Crunchbase Daily.