From a $5M seed to a $100M debt facility for a sovereign AI company

The company had real assets - a credible path to power, a founding team that could actually build, and an early commercial pipeline - but no story that made an investor lean in before slide five.

When we joined, the question wasn’t “is this a good company?” It was “why should anyone fund this data center rather than the other forty in their inbox?”

Data center investors don’t buy the AI story. They buy five things, in roughly this order.

1. Time to power.

   Power availability - not capital - is now the principal driver of investment decisions. “Will-serve” letters do not equal powered land; what matters is a contract for transmission capacity by a date certain. A site with signed tenant interest but no credible path to energization is not a data center investment - it is a land option with development risk. We made time-to-power the spine of the deck: grid connection status per site, dated milestones, and a clear argument for why the company reaches energised megawatts faster than anyone else in its markets.

   

2. Contracted demand.

   The entire neocloud capital structure can be summarized in one rule: offtake gets you infrastructure financing. Lenders underwrite the offtake counterparty far more than the hardware. We built the traction section to show a pipeline that compounds - counterparties, contract tenor, and how each signed megawatt pulls forward the next site. Not “we have interest,” but “here is who, for how long, and what it funds.”

   

3. Site-level economics.

   AI has raised underwriting standards: investors must now assess rack density, cooling requirements, customer credit quality, equipment cycles, and power intensity. Each site - Italy, Switzerland - got its own model: power price, capex per megawatt, ramp, utilisation, rolled up into the group model. It turned “trust us” into “here’s the sensitivity table.”

   

4. A capital structure that matches the asset.

   Equity for the platform and the pipeline; debt for the sites and the fleet. In GPU-backed raises the chips and their contracts are sold into a bankruptcy-remote special-purpose entity so that if the operating company fails, the collateral is still there and still reachable. Showing the debt path from the seed deck onward made the equity ask smaller and the return story cleaner - and it laid the groundwork for the $100M facility that followed.

5. Honesty about the asset.

   Rental rates for H100-class capacity fell from roughly $7 to $10 per hour in early 2024 to $2 to $4 by late 2025. A deck that pretends GPUs don’t depreciate gets binned; one that models refresh cycles and shows the site economics still work is the one that gets the second meeting.

   

Sovereignty is a moat only when paired with demand and speed. Private capital is not the crucial constraint on Europe’s AI compute build-out; speed-to-operation and grid access are more significant limitations. The companies that win the sovereign race are the ones that can build, not the ones that can lobby.