Jasper raised $125 million at a $1.5 billion valuation in October 2022. The product was a friendly interface on top of OpenAI’s GPT-3, sold to marketers and content creators who didn’t want to write prompts directly. Six weeks after that round closed, OpenAI launched ChatGPT, giving the world a friendly interface on top of GPT-3 for free. Within nine months, Jasper had laid off staff, cut its internal valuation, and started pivoting away from the broad creator audience it had built the company around. The CEO did not pretend it was anything else. ChatGPT, he wrote, “remade the industry’s landscape.”
When your supplier launches your product for free, you find out what your moat actually was.
That story repeats. Copy.ai, Writesonic, dozens of “AI for X” tools that raised on the promise of a defensible niche on top of someone else’s API. Most are now smaller, narrower, or absorbed.
There are two ways the supplier kills you. They can offer your product directly. Or they can drop the price of the input until your markup looks ridiculous.
In January 2025, DeepSeek, a Chinese lab, released R1: a model close to OpenAI’s frontier reasoning at a fraction of the cost. OpenAI’s CEO publicly acknowledged R1 ran 20 to 50 times cheaper than the equivalent OpenAI model. Nvidia’s stock crashed. What followed was the loudest year of API price cuts the industry had seen, with DeepSeek, Anthropic, Google, and OpenAI all cutting repeatedly through 2025.
In theory, cheaper APIs should help any business built on top of them. In practice, the opposite happened. When customers can see roughly what an API call costs, they won’t pay a 10x markup for a thin layer on top. The wrapper’s input cost dropped. Its pricing power dropped faster. And as the underlying models got more capable, “I’ll just use ChatGPT directly” became a more viable answer to whatever the wrapper was solving. Margins, often 20 to 40% on a good day, compressed toward zero.
Picture the partner meeting at a mid-sized fund somewhere through this. Half the AI portfolio is wrappers. The discussion is no longer “how do we save these companies?” It is something more uncomfortable: what did we actually invest in?
The model is the product
Whatever your product depends on most heavily, that is what you are actually selling. A startup with rare medical imaging data and a generic AI model is selling the data. The model is interchangeable. They could swap it next month and the customer wouldn’t notice. But a startup with public data and a clever prompt is selling the model itself, dressed up. The model is the product, even if the founder doesn’t want to call it that.
If the model is your product and you rent it, you are renting your product. If the rent goes up, your margin goes down. If the landlord launches a competing tenant, you are competing with your own supplier.
That is platform risk, and it is the most under-priced item in the AI startup category today.
The compute wrapper risk
The AI wrapper conversation stops at the model layer. It shouldn’t.
A startup that “owns its infrastructure” by renting H100s from AWS or Azure is a compute wrapper. The cloud provider rents from Nvidia. Nvidia depends on TSMC. At every layer, someone above can change the price, the terms, or whether you get any chips at all, and the layer below has to absorb the hit.
This is not binary. There is a spectrum. At one end, a pure wrapper that adds a logo and a login screen. At the other, a company that has built so much around the model (proprietary data, specialized workflows, deep customer integration) that the model becomes one swappable input among many. The compute layer has the same spectrum. Renting compute is not a death sentence. The question is how exposed you are, and what you’ve built around it that would survive if the rental terms changed tomorrow.
Most pitch decks today don’t ask this question at any layer. They should ask it at every layer.
The math
A single H100-class server, the kind of machine you’d use to run a serious open-source model in production, costs $15,000 to $40,000. Running a 7B-parameter model on that hardware at 70% utilization costs roughly $10,300 a year all-in. The same workload on a hosted API runs into the hundreds of thousands of dollars at any meaningful volume.
Lenovo’s 2026 analysis puts the on-premises advantage at up to 18x cheaper per million tokens than hosted APIs for steady inference, with breakeven in under four months for high-utilization workloads.
That kind of saving used to be a finance optimization. In an AI business, it’s the difference between having gross margin and not having it.
And the rented version is getting more expensive, not less. In November 2025, Nvidia’s CFO told investors the clouds are sold out and every GPU they have shipped is in use. Lead times for data-center GPUs run 36 to 52 weeks. The price of the memory chips inside these servers nearly tripled in three months at the end of 2025. You cannot assume you’ll be able to rent your way out of a capacity problem in 2026.
But owning a GPU is not a moat either
All of which makes a strong case for owning your compute. Just don’t confuse that with having a moat.
There are companies right now raising rounds on the strength of “we own our compute,” and that alone is not a defensible position. A competitor with a credit card and an API key can match you in a week.
The moat is never the hardware. The moat is what the hardware lets you build that others cannot easily copy:
- A model trained on data nobody else can legally or practically get
- A cost per request so low that you can serve customers your venture-funded competitors cannot serve profitably
- Speed or reliability that becomes a feature your customer cannot get from a generic API
- A regulatory position (EU data residency, healthcare compliance, defense clearance) where the architecture is the qualification
Hardware could be a requirement for these moats. A data moat does not necessarily need owned hardware. A cost moat almost certainly does.
For most SaaS companies, the real leverage point isn’t hardware. It’s data. Specifically: data the company already has through its product, that competitors cannot legally or practically obtain. Building AI features on top of that data is the difference between adding a feature anyone can copy and creating something that gets harder to replicate every month it runs. This is the work my team at Wise Minds does with SaaS scale-ups: turning existing data and workflows into AI features that compound. If your product has real users and real data and you’re feeling pressure to “do something with AI,” the first question is not which model to use. It’s which part of your data nobody else has, and what AI feature would not exist without it.
Why this matters more in Europe
The compute scarcity story is global. A founder in San Francisco faces the same lead times as a founder in Berlin.
What’s specifically European is the sovereignty layer on top.
If your European startup processes EU customer data through a model hosted by a US provider, on chips whose export is controlled by US policy, you have three points of foreign control in your stack before you even open the GDPR conversation. The European Commission is funding a network of AI Gigafactories, with each planned to host around 100,000 state-of-the-art AI chips. The largest US AI labs are operating at orders of magnitude more compute than this combined. The gap matters less for any specific startup than what it signals: Europe is competing for compute capacity it does not yet have, with suppliers it does not control.
For European founders, this is not abstract. It is the difference between winning a regulated enterprise customer and getting a polite “no” because the architecture cannot pass their compliance review.
The new diligence question
This is supplier power. Porter wrote about it decades ago. When suppliers are concentrated, when their inputs are critical, when switching is hard, when they can step into your market and become your competitor, margins compress and strategy gets dictated from outside the building.
That is the AI stack right now. A handful of model providers. A handful of cloud providers. One dominant chip vendor. OpenAI launching Custom GPTs is the textbook case of a supplier stepping into its customers’ market.
The diligence question is just supplier power, named honestly:
What happens to this company if the platform it depends on changes the rules?
If the answer is “we’d be in serious trouble within 90 days,” you are looking at a wrapper, no matter which layer it sits on.
For founders, the same question, asked of yourself:
- List every outside dependency in your stack: model provider, compute, hosting, distribution, data access.
- For each one, write: “If [dependency] doubled their price tomorrow, we would…”
- For each one, write: “If [dependency] launched a competing product tomorrow, we would…”
- The answers without good responses are not risks. They are timers.
In short
The same thing that made these companies possible (cheap, accessible intelligence sitting one API call away) is what made them disposable. The companies still standing in 2028 will be the ones who used the wrapper era to build something the supplier cannot copy and cannot undercut.
The investors who figure that out first will rewrite their thesis in the next two quarters. The founders who figure it out first will still have a company.
