We raise to build compute, not to buy market share
Capital goes into GPU inventory, colocation and power, underwritten against rental revenue we can already forecast. Alongside that, we hold minority positions in operators outside of infrastructure entirely.
Hardware is the asset, rental income is the yield
A GPU depreciates on a known schedule and earns rental income from the day it's racked. We underwrite each purchase against contracted and forecast utilisation before we buy it, not after; the fleet is sized to demand we can already see, plus a margin for growth we're confident in.
That's a different bet than most compute-adjacent raises, which fund a cloud bill paid to someone else. Every dollar here converts into equipment we hold the title to, in a site we have a direct power contract for. If the company stopped raising tomorrow, the fleet keeps earning.
How the yield actually works
Every tranche is sized to hardware, not to a target multiple. The math is simple enough to check yourself, which is the point.
A rolling close, not a fixed vintage
We don't run a single annual round. Capital comes in as tranches, each sized to a specific hardware purchase and underwritten against the rental revenue that purchase is already forecast to earn.
That means the terms on offer at any given time reflect exactly what the next tranche is funding; there's no blended cap table story to explain, and no pressure to deploy capital faster than the fleet can absorb it responsibly.
How the capital is protected
Infrastructure investing lives or dies on what happens if a raise underperforms. We underwrite conservatively and report often, so the answer is never a surprise.
We invest across industries we don't operate in
Alongside infrastructure, we hold minority positions in companies well outside it. The common thread is a business past its first million in revenue with one function it can't build alone, usually engineering or data, occasionally both.
We supply capital and, where it helps, the same engineering team that builds our internal tools. It's a small program run by the same twenty-one people as everything else here, so we're selective about where we say yes.